Research ArticleDeep Learning Hash for Wireless Multimedia ImageContent Security
Yu Zheng 1 Jiezhong Zhu1 Wei Fang1 and Lian-Hua Chi2
1School of Computer amp Software Jiangsu Engineering Center of Network MonitoringNanjing University of Information Science amp Technology Nanjing 210044 Jiangsu China2Department of Computer Science and Information Technology La Trobe University VIC 3086 Australia
Correspondence should be addressed to Yu Zheng yzhengnuisteducn
Received 24 July 2018 Accepted 30 August 2018 Published 25 September 2018
Academic Editor Weizhi Meng
Copyright copy 2018 Yu Zheng et alThis is an open access article distributed under the Creative CommonsAttribution License whichpermits unrestricted use distribution and reproduction in any medium provided the original work is properly cited
With the explosive growth of the wireless multimedia data on the wireless Internet a large number of illegal images have beenwidely disseminated in wireless networks which seriously endangers the content security of wireless networks However how toidentify and classify illegal images quickly accurately and in real time is a key challenge for wireless multimedia networks To avoidillegal images circulating on the Internet each image needs to be detected extracted features and compared with the image in thefeature library to verify the legitimacy of the image An improved image deep learning hash (IDLH)method to learn compact binarycodes for image search is proposed in this paper Specifically there are three major processes of IDLH the feature extraction deepsecondary search and image classification IDLH performs image retrieval by the deep neural networks (DNN) as well as imageclassification with the binary hash codes Different from other deep learning-hash methods that often entail heavy computationsby using a conventional classifier exemplified by K nearest neighbor (K-NN) and support vector machines (SVM) our methodlearns classifiers using binary hash codes which can be learned synchronously in training Finally comprehensive experimentsare conducted to evaluate IDLH method by using CIFAR-10 and Caltech 256 image library datasets and the results show that theretrieval performance of IDLH method can effectively identify illegal images
1 Introduction
With the rapid development of multimedia technology theretrieval of wireless multimedia data has become very conve-nient and easy The security of multimedia data has attractedmore and more attention from researchers Multimediadata content security belongs to the branch of informationsecurity and requires direct understanding and analysis ofthe information content transmitted in the network Judgingrapidly from the massive information filtering and mon-itoring the abnormal information in the network are thekey to ensure the security of the wireless network contentMeanwhile a large number of illegal images are disseminatedin the network which seriously endangers the security ofnetwork content So it is very important practical significanceto research the content security-oriented image recognitiontechnology and identify and supervise illegal image informa-tion in the network Although there are a lot of image retrieval
methods currently due to the various disadvantages such asthe low expression ability of image feature high dimensionof feature and low precision of image retrieval the retrievalresults are not always effective
How to retrieve the large-scale image resources quicklyand effectively meet the needs is urgently to solve Since thehash-based learning algorithm can effectively preserve thesimilarity between the original feature spaces and the hashcode spaces more and more scholars have drawn attentionto it Particularly learning deep hash algorithm has greatlyimproved the retrieval performance
Besides the widely used text-based search methodscontent-based image retrieval (CBIR) has attracted wide-spread attention ofmore andmore scholars in the past decade[1] As we all know the nearest neighbor search algorithm isan effectivemethod for searching for similar data samples [2]We should not only consider the scalability issue but alsoconsider most practical large-scale applications which are
HindawiSecurity and Communication NetworksVolume 2018 Article ID 8172725 13 pageshttpsdoiorg10115520188172725
2 Security and Communication Networks
affected by dimensional disasters [3]Thehigh-level encodingcomplexity prevents widespread adoption in real-timemulti-media systems [4] One of the many practical applicationsthe approximate nearest neighbors (ANN) is very efficientHowever the method requires huge storage costs and hardto handling high-dimensional data Hence quantizationtechniques have been proposed to encode high-dimensionaldata vectors recently Due to the fact that hashing-basedANN search techniques can reduce the storage via storingthe compact binary codes hash technology is widely usedin computer vision machine learning information retrievaland other related fields in view of the retrieval of large-scaledata The goal of hash is to turn high-dimensional data intothe low-dimensional compact binary codes [5] For example119889 dimension data is turned into 119903 dimension (dgtgtr) andgenerally 119903 dimension data is between dozens of bits andhundreds of bits After turning high-dimensional data intohash code we calculate the distance or similarity betweenthe data rapidly Because of the high efficiency of binary hashcode in Hamming distance calculation and the advantages ofstorage space hash code is very efficient in large-scale imageretrieval
One of the most famous jobs in hash and the mostapplied job in the industry was locality sensitive hashing(LSH) [3] which was put forward by Gionis Indyk andMotwani in 1999[1] LSH is one of the most popular dataindependent methods which generates hash functions byrandom projections [6] In addition to traditional Euclideandistance LSH has been generalized to accommodate otherdistance and similaritymeasures such as p-normdistance [3]Mahalanobis metric [7] and kernel similarity [8] And L QiX ZhangW Dou andQ Ni put forward another applicationdirection of LSH in 2017 [9] They proposed a privacy-preserving and scalable service recommendation approachbased on distributed LSH (SerRecdistri-LSH)The purpose ofthis application is to use SerRecdistri-LSH method to handleservice recommendation in distributed cloud environment
A disadvantage of the LSH family is that LSH usuallyneeds long bit length (ge 1000) to achieve both high precisionand recallThismay leads to a huge storage overhead and thuslimits the sale at which an LSH algorithm may be appliedIn order to achieve the desired search accuracy LSH oftenrequires the long hash codes thereby reducing the recall rateThis problem can be alleviated by using multiple hash tablesbut it greatly increases storage costs and query time [10]
So learning-based data-dependent hashing methodshave become increasingly popular because of the benefitthat learned compact binary codes can effectively and highlyefficiently index and organize massive data [3] The goal ofthe data-dependent hashing methods is to generate shorthash codes (typically le 200) using available training dataVarious hashing algorithms have been proposed in theliterature of which a large category focuses on linear hashingalgorithms which learn a set of hyperplanes as linear hashfunctions The representative algorithms in this categoryinclude unsupervised PCA Hashing [11] Iterative Quanti-zation (ITQ) [12] and Isotropic Hashing [13] and super-vised Minimal Loss Hashing (MLH) [14] SemisupervisedHashing (SSH) [11] Supervised Discrete Hashing (SDH) [6]
Ranking-Based Supervised Hashing [15] FastHash [16] etcA bilinear form of hash functions was introduced by [1718]
As an extension of linear hash functions a variety ofalgorithms have been proposed to generate nonlinear hashfunctions in a kernel space including Binary ReconstructiveEmbedding (BRE) [10] RandomMaximumMargin Hashing(RMMH) [14] Kernel-Based SupervisedHashing (KSH) [12]and the kernel variant of ITQ [13] In parallel harnessingnonlinear manifold structures has been shown to be effectivein producing compact neighborhood-preserving hash codesThe early algorithm in this fashion is Spectral Hashing (SH)[19] which produces hash codes through solving a contin-uously relaxed mathematical program similar to LaplacianEigenmaps More recently Anchor Graph Hashing (AGH)[20] leveraged anchor graphs for solving the eigenfunctionsof the resulting graph Laplacians making hash code trainingand out-of-sample extension to novel data both tractableand efficient for large-scale datasets Shen et al [6 21 22]proposed a general induction
The rest of this paper is organized as follows Sec-tion 2 briefly reviews the related works about deep-learning-hashing And we present framework and steps of theimproved deep learning-based hash IDLH algorithm indetail in Section 3 Section 4 evaluates the effectiveness ofthe IDLH through a series of contrastive experiments andcarefully analyzes the experimental results followed by theconclusion in Section 5
2 Related Work
Recently deep-learning-hashing as a popular research topichas drawn increasing attention and research efforts in infor-mation retrieval computer vision and machine learningSemantic hashing is recognized as the earliest starting deeplearning hash [23]This method establishes a deep generativemodel to discover hidden binary features Such a deep modelis made as a stack of restricted Boltzmann machines (RMBs)[24] After learning a multilayer RBM by pretraining andfine-tuning the document collection the hash code of anydocument is obtained by simply thresholding the deepestoutput Such a hash code provided by deep RBM is shownto maintain a semantically similar relationship of inputdocuments into code space where each hash code (or hashkey) is used as a memory address to locate the correspondingdocument In this way semantically similar documents aremapped to adjacent memory addresses enabling efficientsearching through hash table lookups In order to improvethe performance of deep RBMS a supervised version wasproposed in [25] and the idea of nonlinear neighborhoodcomponent analysis (NCA) embedding in [26] was adoptedThe supervised information is derived from training a givenneighbornonneighbor relationship between samples Thenbased on the depth RBM the objective function of the NCAis optimized so that the depth RBM yields a hash codeNote that supervised deep RBMS can be applied to widedata fields other than text data In [25] the depth RBMs aresupervised using a Gaussian distribution and the models of
Security and Communication Networks 3
the visible units in the first layer are successfully applied tothe processing of massive image data
In [27] a deep neural network was developed to learnmultilevel nonlinear transformations mapping the originalimage to a compact binary hash code to support large-scaleimage retrieval for learning binary image representationsA deep hashing model is established under the three-layerconstraint on the deep network (1) the reconstruction errorbetween the original real-valued image feature vector andthe resulting binary code is minimized (2) each bit of thebinary code has a balance (3) all bits are independent of eachother Similar constraints were used in previous unsupervisedhash or binary codingmethods such as iterative quantization(ITQ) [28] In [27] a supervised version is called superviseddepth hash 3 in which a discriminative item containing twopairs of supervised information is added to the objectivefunction of the deep hash model
It is worth noting that in the training of sparse neuralnetworks in addition to the sparse similarity maintaininghash depth hash and supervised depth hash the pretrainingphase is not included Instead the hash codes are learnedfrom scratch using a set of training data However nopretraining can make the general hash code less efficient Inparticular the sparse similarity keep hash method is foundto be inferior to the existing supervised hashing methodie kernel-based supervised hashing (KSH) [29] in terms ofsearch accuracy on some image datasets [30] the deep hashmethod and its supervised version are slightly better than ITQand its supervised version CCA+ ITQ respectively [31] Notethat KSH ITQ and CCA+ITQ develop a shallow learningframework
One of the main purposes of deep learning is to learnthe robust and powerful representation of complex data Itis very natural to use deep learning to explore compact hashcodes which can be thought of as binary representation ofthe data The deployed CNN consists of three convolutionpools including rectified linear activation maximum poolmerging and local contrast normalization a standard fullyconnected layer and an output layer with softmax activationfunctions In [32] a newmethod called deep semantic sortinghashing is proposed to learn hash values therebymaintainingmultilevel semantic similarity between multilabel imagesThis method is combined with convolutional neural networkhashing method taking image pixels as input training depthCNN and jointly learning image feature representation andhash value by thismethodThe deployed CNN consists of fiveconvolution-pooling layers two fully connected layers and ahash layer (ie output layer)
Indexing massive amounts of multimedia data such asimages and videos is a natural application based on learninghashing In particular due to the well-known semanticdivide hashing methods have been extensively studied forimage search and retrieval as well as mobile product search[33 34] Although hashing techniques have been appliedto the active learning framework to cope with big dataapplications these hash algorithms are rarely applied toimage classification For the image recognition problem withmany categories the computational and memory overheadprimarily stems from the large number of classifiers to be
Table 1 A huge number of parameters need to be learned andstored with different datasets Considering a classification task withC different categories and D-dimensional feature representationeven the simplest linear models are comprised of DtimesC parameters
learned Table 1 show that the complexities can be high atthe stages of both training and deploying these classifiersInspired by [35 36] we proposed a combination classifierswith DNN and binarizing classifiers to solve some classicsecurity calculation problems [37ndash39] Different from otherdeep learning-hash methods that often entails heavy compu-tations by using a conventional classifier exemplified by K-NNand SVM ourmethod learns classifiers using binary hashcodes which are simultaneously learned from the trainingdata The combination classifiers can provide both imagefeatures and accelerate image classification and thus it makethe large-scale image recognition faster The advantages ofthe extending hashing techniques from fast image search toimage classification inspire us to apply them to deep learninghash framework
With the development of Internet technology the spread-ing form of illegal information on the Internet is changinggradually Traditionally the form of communication basedon word description has been transformed into a diversifiedformof communication based on video and imageThereforethe original keyword blocking web content grading blacklistrestriction and other filtering methods have not been able toblock illegal information dissemination At present networkillegal image recognition is mainly divided into the followingcategories
(1) Erotic Image Recognition The identification of onlinepornographic imagesmainly includesmethods based on skincolor feature extraction detection and judgment andmethodsbased on limb judgment The limb state is determinedby extracting the divided skin color information and theconnection feature of the human body posture and furtherdetermining whether the image transmitted on the networkcontains pornographic content
(2) Images Involving State Secrets or Military Secrets Therecognition of secret-related images mainly involvessteganography which hides secret information in imagevideo and other carriers for secret transmission At presentthe high-order statistical features based on modeling thecomplex correlation of image neighborhood become themainstream features in the field of steganalysis SRM(Spatial Rich Model) PSRM (Projection Speciation RichModel) and other models are based on such high-orderhigh-dimensional features and have achieved good detectionresults Steganalysis based on depth learning is a hotspot in
4 Security and Communication Networks
the field of information hiding in order to identify the illegalsecret-related images accurately and quickly
(3) Images Containing Antihuman Content such as TerroristViolenceThe identification of such images is mainly based onimage contrast techniques Image comparison includes tech-niques such as image feature extraction high-dimensionalspatial feature index establishment and similarity measureIt is a very worthwhile to study how to quickly comparethe massive network images to the illegal target images sothat the recall rate and the precision rate can be taken intoaccount
3 The Proposed Method
In this section we will present the notations as summarizedin Table 2 firstly The concept of deep learning stems fromthe field of artificial neural networks Deep learning is deepneural network learning and is a learning structure withmultiple hidden layers In the process of deep learningthe network is trained layer by layer Each layer of thelearning network extracts certain features and informationand takes the training result as a deeper input Finallythe entire network is fine-tuned with a top-down algo-rithm
Through deep learning complex function expressionscan be learned thereby completing the concept of high-levelabstraction from the underlying information It has beenwidely used in language understanding target recognitionand speech perception
Figure 1 shows that the proposed framework IDLHincludes three components The first component is prepro-cessing layer on the image dataset The second component isthe training self-coding network with a layer-by-layer greedylearning algorithm to obtain the feature expression functionof the image The third is hash layer which retrieves imagessimilar to the query image with compact binary codes andcategorizes the query one by the majority semantic labelwithin the hashing bucket
31 Preprocessing Since the depth learning algorithm usedin this paper is an unsupervised learning algorithm it canautomatically learn the deep features of the image fromthe original pixel information of the image Therefore theoriginal pixel value of the whole image can be directly used asinput data for the deep learning model In order to facilitatethe training of the network it is necessary to preprocess theimage Through preprocessing the image is simply scaledsample-by-sample mean-value reduction and whitening isprocessed to reduce the redundant information in the imageand facilitate the deep learning network for training andcalculation Preprocessing can be further grouped into threesuboperations
(1) Normalization Normalization can prevent neuron outputsaturation caused by excessive net input absolute value Weuse the sigmoid function to do normalization as shown asfollows
Table 2 Summary of notations
Symbol Definition119909119894 the gray value of the image pixels120583(119894) mean pixels
U an arbitrary orthogonal matrix and defines in theZCA whitening
J(Wb) the quantization loss between the learned binaryvalues and the real values
120582 a weight attenuation parameter119878119894 119894119905ℎ seta b two thresholdsD dimensionality of data pointsM K number of candidate imagesE the objective function
119909119894 = 1(1 + 119890minus119909119894) (1)
119909119894 is the gray value of the image pixels
(2) Pointwise Mean Reduction This process mainly is to getrid of the redundant information of the image the meanvalue is eliminated for each point of the image the averagebrightness of the image is removed and the DC componentof the data is eliminated Assuming that 119909(119894) isin 119877119899 is the grayvalue of each pixel of image I we use formulae (2) and (3) tozero-mean image
120583(119894) = 1119899119899sum119895=1
119909(119894)119895 (2)
119909(119894)119895 = 119909(119894)119895 minus 120583(119894) (3)
(3) Whitening Whitening is an important pretreatmentprocess its purpose is to reduce the redundancy of input dataso that the whitened input data has the following properties(i) low correlation between features (ii) all features having thesame variance then the formula is as formula (4)
In formula (4) the rotation matrix of 119909119903119900119905119894 is 119880119879119909119894Generally when 119909 is in interval [-11] 120576 asymp 10minus5
32 Training Stack Sparse Self-Encoding Network Stack self-coding neural network has strong expressive ability whichmainly benefits from its hierarchical feature representationThrough one level of feature learning we can learn thehierarchical structure between features Stack self-encodingneural network is a neural network model composed ofmultilayer sparse self-encoder that is the output of theformer self-encoder as the input of the latter self-encoder
In the training the original input 119909(119896) is used as inputto train the first self-encoded neural network At this pointfor each training sample 119909(119896) the output ℎ(119896)1 of the hidden
Security and Communication Networks 5
Hash Layer 2
Hamming Retrieval
Query Result
Weights
Codes
Training Dataset
Feature expression
function
Image Preprocess Layer Feature
expression function
Hash Layer 1
Binary Weights
Binary Codes
Feature migration
Query Image
Figure 1 Deep learning-hash retrieval framework IDLH consists of three main components (preprocessing layer deep neural networklayer and hash layer) The object of the first layer is simply scaled sample-by-sample mean-value reduction and whitening In the secondcomponent we develop a deep neural network to obtain the feature expression function of the image And the classifier weights and featurebinary codes are simultaneously learned in the last component-hash layer
layer can be obtained and the output of the hidden layer canbe used as the input of the second self-encoder to continuetraining the second self-encoder Then the output ℎ(119896)2 of thesecond hidden layer of the self-encoder can be obtained Theoutputℎ(119896)1 of the first hidden layer of the self-encoder is calleda first-order feature and the output ℎ(119896)2 of the second hiddenlayer of the self-encoder is called a second-order feature Inorder to classify the two-order feature ℎ(119896)2 can be used as theinput of Softmax regression
Figure 2 shows the flowchart of the proposed methodAnd there are mainly three processes (supervised pretrain-ing rough image retrieval and accurate image retrieval)The object of the first process is to transform the high-dimensional feature vector into a low-dimensional compacttwo value codes through hash function In the secondprocedure we pick out M candidate images by calculatingHamming distance In the third process we calculate theEuclidean distance between the candidate image and theimage to be retrieved and accurately extract K images fromthe M candidate images
Figure 3 shows a block diagram of a self-encoding neuralnetwork The stacking self-encoding network contains 3hidden layers (feature layers) The input layer inputs theoriginal data i into the first layer of the feature layer theoutput result of the former layer serves as the input of thenext layer and the output of the third layer serves as thefeature expression of the image In our method it is also usedas the input of the hash classifier and it is possible to usethe characteristics of the STD neural network to classify thefeatures
By using the matrix representation of the binary codesvectors and the output of the 3th layer of the network we usethe gradient descent method to solve the neural network
For a single sample (x y) the cost function is as shown in
(12 10038171003817100381710038171003817ℎ119882119887 (119909(119894)) minus 119910(119894)100381710038171003817100381710038172)]
+ 1205822119899119897minus1sum119897=1
119904119897sum119894=1
119904119897+1sum119895=1
(119882(119897)119895119894 )2
(6)
The first term 119869(119882 119887) represents the mean variance termThe second term aims to prevent the data from overfitting byreducing the magnitude of the weight 120582 is a weight attenua-tion parameter It is used to balance the relative importance ofmean square deviation terms and weight attenuation termsOur purpose is to minimize the quantization loss 119869(119882 119887)between the learned binary values and the real values of aninput image according to parameters119882 and 11988733 Hash Algorithm Retrieval The image retrieval methodbased on hash algorithmmaps the high-dimensional contentfeatures of images into Hamming space (binary space) andgenerates a low-dimensional hash sequence to represent apicture This method reduces the requirement of computermemory space for image retrieval system improves theretrieval speed and better adapts to the requirements of massimage retrieval
Inspired by [6 8] we use a set of hash functions to hashdata into different buckets After we do some hash mappingon the original image feature data we hope that the originaltwo adjacent feature data can be hash into the same bucketwith the same bucket number And then after hash mappingof all the data in the original feature set we can get a hashtable These original feature data sets are scattered into hashtable buckets and the data belonging to the same bucket isprobably adjacent to the original data However there is also
6 Security and Communication Networks
StartTwo valued by the
Sigmod function in Hash layer
Building hash pool and hash bucket
Calculation of Hamming distance
Select the top m candidate images
Calculation of Euclidean distance
Select the top kimages from m
candidate imagesas retrieval result
End
Supervised pre-training
Rough image retrieval
If itgt05Yes
1
No
0
If it less than a threshold
Yes
No
Accurate image retrieval
The featureof the image
to be retrieved
Self-learning network
The featureof the
training images
Binary processing
Binary codes for trainingimage
Binary codes for retrieving
image
Figure 2 Deep learning-hash retrieval flowchart IDLHmainly includes three processes (supervised pretraining rough image retrieval andaccurate image retrieval)
+1+1 +1
+1
Layer 0Input layer
Layer 1 Layer 2 Layer 3 Hash Layer
x1
x2
x3
x4
x5
ℎ(1)1
ℎ(1)2
ℎ(1)3
ℎ(1)4
ℎ(2)1
ℎ(2)2
ℎ(2)3
ℎ(2)4
ℎ(3)1
ℎ(3)2
ℎ(3)3
Output ℎ(3)k
from layer 3 is usedas binary code with bk = f(ℎ(3)
k)
Figure 3 Self-learning network based on stack self-encoding network The neurons labeled xi is the input of the neural network and ldquo+ 1rdquoare the offset nodes (intercept entries) of the neural network The layer 0 is the input layer of neural network and layer 3 is the output layerof neural network The middle layers of layer 0 layer to layer 3 are the hidden layers of neural network
a small probability in events that is the nonadjacent data ishash to the same barrel Set the hash function as the follows
Here119908119896 is the projection vector and 119887119896 is the correspond-ing intercept The code value generated by formula (7) isminus1 1 and we use the following formula to convert it intotwo value codes
119910119896 = 12 (1 + ℎ119896 (119909)) (8)
Given a sample point isin 119877119863 we can compute a K-bitbinary code 119910 for 119909 with formula (9) The hash functionperforms the mapping as ℎ119896 119877119863 997888rarr 119861
Here 119883 = 119909119899119873119899=1 isin 119877119863times119873 is the feature data matrix withpoints as columns Such a binary encoding process can also beviewed as mapping the original data point to a binary valuedspace namely Hamming space
34 SimilarityMeasure After obtaining the binary hash codeof the image it is necessary tomeasure similarity between theretrieved image and the library image in the Hamming spaceThe smaller the Hamming distance is the closer distancebetween the two data is and the degree of similarity is higherotherwise the two data similarity is lower
Here oplus is an XOR operation The two sets of 119910119894 and 119910119895represent the hash code of the search image feature and theimage library is mapped through the hash function The newimage features learned by the stack self-encoding network aregenerated by the hash function The storage structure of theimage feature vectors is shown in Table 3
As can be seen from Table 3 the hash code of the imageis related to the image ID and the image name one by one Inthe process of searching the image feature vector is obtainedthrough deep learning by a hash function the original data ismapped into a newdata space and a corresponding hash codeis obtained The hash code is used to calculate the Hammingdistance in the Hamming space as a measure of similaritybetween images Finally the storage structure of the imagefeature vector is used to find the corresponding image ID ofthe hash code and the output retrieval result is output to theuser
35 Image Secondary Search In the first-level search phasethe features learned from the deep learning network aremapped into the Hamming space using the hash function Inthe similarity measurement phase the traditional Euclideandistance is abandoned Measure the similarity betweenimages by comparing the Hamming distance between theimage features of the query image and the image of thelibrary image In order to further improve the accuracy ofretrieval without affecting the real-time performance we canretrieve the image by the second level retrieval These stepsare described in detail as follows After one level retrievalwe choose the K images with the most similarity in thefirst-level retrieval result and then calculate the Euclideandistance between the original feature vector of the K imagesand the original feature vector of the query imageThe resultsobtained as the similarity measure of the images and outputthe retrieval result that has been ranked from the high andlow with the similarity distance
Although theHash algorithmmaps the high-dimensionalfeature vectors of the image into a hash-coded form theproblem of ldquodimensional disastersrdquo is solved and the retrievalefficiency is greatly accelerated However when the similaritycomparison is performed the Hamming distances of theimage features are simply compared using the results of theprimary search and occasionally undesirable results maystill appear on the search results If we want to increasethe accuracy of the search we must increase the hash codelength However excessively long codes will increase theamount of calculations increase the memory burden andreduce the real-time nature of retrieval failing to achieve
the goal of reducing the size of data In order to solve thisproblem keep the retrieval efficiency and further improvethe retrieval accuracy we propose a search strategy forsecondary retrieval the specific steps of which are as follows
Step 1 Through the first-level search in the Hamming spacethe similarity degree of the images is sorted and the top Ksorting images are selected
Step 2 For the 119896 images in Step 1 calculate the Euclideandistance one by one from its original image feature vector tothe image feature vector of the query image
Step 3 The Euclidean distance calculated in Step 2 is sortedThe smaller the calculated value is the higher similaritybetween images is and the similarity is sorted from high tolow and output as the final search result
In the second search it is necessary to pay attention tothe selection of the 119870 value although the larger the 119870 valueis the better the search effect is but accordingly the longerthe time is consumed Therefore it is necessary to combinevarious factors to select the appropriate119870 value
4 Experimental and Performance Analysis
In this section we thoroughly compare the proposedapproach with the improved deep learning hash retrievalmethods on several benchmark datasets Through a series ofexperiments the effectiveness and feasibility of the proposedalgorithm are verified
41 Database Two mostly used databases in the recent deeplearning hash works are taken into our evaluation Thetwo image libraries are derived from the CIFAR-10[11] coreexperimental image library dataset and theCaltech 256 imagelibrary dataset
CIFAR-10 dataset contains 10 object categories and eachclass consists of 6000 images resulting in a total of 60000images The dataset is split into training and test sets whichare averagely divided into 10 object classes The Caltech 256image library dataset contains 29780 color images which aregrouped intro 256 classes
First test the CIFAR-10 image dataset There are a totalof 50000 training samples which are used for training on thedeep learning networkThe remaining 10000 images are usedas test samples And then we randomly select 50 images fromdatabase as the query images For theHidden Image Retrievalalgorithm based on deep learning mentioned in this paperthe image pixel data is directly used as input while for otheralgorithms the 512-dimensional GIST feature is used as thefeature expression of the image Note quantization all imagesinto 32lowast32 sizes before experiment
For the Caltech 256 image a total of 256 classes areincluded and each class contains at least 70 images There-fore 70 images of each class a total of 17920 images arerandomly selected and are used as training images Theremaining images are used as test samples In addition all ofthe imagesrsquo size is set to 64lowast64 again when training
8 Security and Communication Networks
42 Evaluation Metrics We measure the performance ofcompared methods using Precision-recall and Average-Retrieval Precision (ARP) curves Precision is the ratio ofthe correct number of images m in the search result to thenumber k of all returned images The formula is as follows
precision = 119898119896 times 100 (12)
Recall is the ratio of the correct number of imagesm in thesearch results to the number g of images in the image libraryThe formula is as follows
recall = 119898119902 times 100 (13)
Assume that the search result of the query image 119894 is119861119894 and 119860 119894 means that the category is the same between thequery image and the return image then the accuracy rate forthe image query result 119875(119894) can be defined by the followingformula
119875 (119894) = |119860 (119894) cap 119861 (119894)||119861 (119894)| (14)
Average-Retrieval Precision (ARP) the average value ofall the images in the same class as the retrieval rate obtainedfrom the retrieval image is defined as follows
Here 119868119863119898 is the category index number of the image119898 isthe category index119873 is the number of imageswhose categoryis 119868119863119898 and 119894119889(119894) is the category index number of the queryimage
43 Performance Analysis In the proposed algorithm IDLHthe length of the hash sequence and the depth of the hiddenlayer in the deep learning network are two key parametersWhen the hash sequence length is small different featurevectors can easily be mapped into the same hash sequence sothe retrieval accuracy is low However if the hash sequence istoo long a large storage space is required and a long time isconsumed which reduces the real-time performance For thenumber of hidden layers the number of layers in the hiddenlayer is too small which is not conducive to learning strongimage features However if the depth of the hidden layer istoo large the difficulty of training is increased In order toverify the effectiveness and feasibility of our algorithm weconducted the following experiments
(1) Results on CIFAR-10 Dataset Figure 4(a) shows the searchresults of the Average-Retrieval Precision using our proposedalgorithm IDLH compared with the LSH algorithm [3] andother three deep learning algorithms the DH algorithm [21]the DeepBit algorithm [40] and the UH-BDNN algorithm[41] on CIFAR-10 dataset with 8 16 32 48 64 96 128and 256 bits Figure 4(b) shows the Precision-recall curveunder 48-bit encoding It can be seen that the algorithm hasa higher precision than the other hashing algorithms with
the same recall rate However the advantage is not obviousand the average accuracy is slightly higher than other hashalgorithms
In order to overcome the above defects we use deeplearning to perform hash mapping on image features per-form hash encoding of different bits on the same featurecalculate the Precision-recall of the search results under thecondition of different coded bits and determine the impactof the encoding length on the retrieval results
As Figure 5 shows with the increase in the number ofcoded bits the Precision-recall is continuously increasingWith the increase in the number of coded bits the imageis better expressed However after the number of coded bitsreaches 64 even if the number of coded bits increases theaverage accuracy rate increases relatively slowly Because theinformation of the tiny image is relatively simple when thenumber of encoded bits reaches 64 bits a relatively goodimage expression has been obtained and the performanceof the algorithm has basically stabilized At this time despiteincrease in the number of encoding bits it is not very helpfulto improve the accuracy rate
In addition we want to test the influence of the numberof hidden layers in the deep learning network on the retrievalresult by changing the number of hidden layers
Figure 6 shows the effect of deep learning networks onexperimental results in the case of different hidden layernumbers
It can be seen that deeper networks do not have muchimprovement in performance which is different from theexpectation that more hidden layers will help learn strongerimage features Since the image library data used in theexperiment is a tiny image library relatively good image char-acteristics can be learned using a deep learning network withfewer layers However if the image library is replaced witha more colorful image the deep neural network can acquiremore detailed image features and the deepening of thelearning network will greatly help the study of image features
(2) Results on Caltech 256 Image Data Set Figure 7(a)shows the results of the Average-Retrieval Precision resultswhen the number of coded bits is different Compared withthe black and white image library the proposed algorithmembodies the advantage of image feature learning and leadsthe Average-Retrieval Precision to other hash retrieval algo-rithms In Figure 7(b) we can also see that the algorithmproposed in this paper has a higher Precision-recall thanother algorithms under the same recall rate and it has bettersearch performance
As shown in Figure 8 as the number of coded bitsincreases the precision rate increases with the same recallrate This feature of the color image library is more pro-nounced than the black and white image library Becausethe color image contains more information more codingis needed to express it and the increase of encoding helpsto learn the features of the image The experimental resultsalso show that the deep learning network has learned moreexcellent image features
In Figure 9 the precision rate is significantly improvedby the increase in the number of hidden layers in the
Security and Communication Networks 9
IDLHLSHDHDeepBitUH-BDNN
16 32 48 64 96 128 2568Bits
0
01
02
03
04
05
06
07
08Av
erag
e Ret
rieva
l Pre
cisio
n
(a) Average-Retrieval Precision
IDLHLSHDHDeepBitUH-BDNN
01 02 03 04 05 06 07 08 09 100Recall
0
01
02
03
04
05
06
07
08
09
Prec
ision
(b) Precision-recall at 48 bits
Figure 4 Five kinds of algorithm retrieval performance comparison on CIFAR-10 dataset
8 bits16 bits32 bits48 bits64 bits128 bits
01 02 03 04 05 06 07 08 09 10Recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
Figure 5 Precision-recall curves with lengths
color image library Caltech 256 This is because theinformation contained in a more colorful image is morecomplex Adding a hidden layer can learn more detailsof the image and help improve the accuracy of thesearch
Next we tested the performance of secondary imageretrieval The value of k in the secondary search is 20 andthe experimental results are shown in Figure 10 As can
Level 1Level 2Level 3Level 4
01 02 03 04 05 06 07 08 09 1000Recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
Figure 6 Precision-recall curves with different code differenthidden layers
be seen from the results secondary retrieval can effectivelyimprove the retrieval accuracy when the number of codedbits is small However with the increase in the number ofencoding bits the results of the secondary search and theaccuracy of the primary search are not much different Thisis because the shorter the hash sequence is the easier thefeature vectors with different original features are mappedto the same hash code In order to make up for the errors
10 Security and Communication Networks
IDLHLSHDHDeepBitUH-BDNN
16 32 48 64 96 128 2568Bits
0
01
02
03
04
05
06
07
08
09
1Av
erag
e-Re
trie
val P
reci
sion
(a) Average-Retrieval Precision
IDLHLSHDHDeepBitUH-BDNN
01 02 03 04 05 06 07 08 09 1000Precision-recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
(b) Precision-recall at 48 bits
Figure 7 Five kinds of algorithm retrieval performance comparison on Caltech 256 set
8 bits16 bits32 bits48 bits64 bits128 bits
01 02 03 04 05 06 07 08 09 10Recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
Figure 8 Precision-recall curves with different code lengths
caused by the short hash code it is necessary to performsecondary searchWhen the number of encoding bits is smalla secondary retrieval method is used in IDLH and the searchaccuracy rate can be improved at the expense of a small searchspeed
Level 1Level 2Level 3Level 4
01 02 03 04 05 06 07 08 09 10Recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
Figure 9 Precision-recall curves with different hidden layers
5 Conclusion
With the rapid development of data storage and digital pro-cess more and more digital information is transformed andtransmitted over the Internet day by day which brings peoplea series of security problems as well as convenience [42] Theresearches on digital image security that is image encryptionimage data hiding and image authentication become moreimportant than ever [38 39] The most essential problem of
Security and Communication Networks 11
First retrievalSecond retrieval
16 32 48 64 96 128 2568Bits
07
08
09
1
Aver
age-
Retr
ieva
l Pre
cisio
n
Figure 10 Average-Retrieval Precision with first and secondretrieval
image recognition is to extract robust features The qualityof feature extraction is directly related to the effect of recog-nition so most of the previous work on image recognitionis spent on artificial design features [43] In recent yearsthe emergence of deep learning technology has changed thestatus of artificial design classification characteristics Deeplearning technology simulates the mechanism of humanvisual system information classification processing from themost primitive image pixels to lower edge features then tothe target components that are combined on the edge andfinally to the whole target depth learning can be combinedby layer by layer The high-level feature is the combination oflow level features From low level to high level features aremore and more abstract and show semantics more and moreFrom the underlying features to the combination of high-levelfeatures it is the depth of learning that is done by itself Itdoes not require manual intervention Compared with thecharacteristics of the artificial design this combination offeatures can be closer to the semantic expression
In terms of illegal image retrieval the traditional recog-nition method should establish a recognition model foreach type of recognition task In the actual application arecognition model needs a recognition server If there aremany identification tasks the cost is too high We used thedeep neural network to recognize the illegal image it onlyneeds to collect the samples of every kind of illegal imageand participate in the training of the deep neural networkFinally a multiclassification recognition model is trainedWhen classifying unknown samples deep neural networkaccounting calculates the probability that the image belongsto each class
We all know that in the image detection process theaccuracy and recall rate are mutually influential Ideally bothmust be high but in general the accuracy is high and the
recall rate is low the recall rate is high and the accuracy islow For image retrieval we need to improve the accuracyunder the condition of guaranteeing the recall rate Forimage disease surveillance and anti-illegal images we need toenhance the recall under the condition of ensuring accuracyTherefore in different application scenarios in order toachieve a balance between accuracy and recall perhaps somegame theory (such as Nash Equilibrium [44 45]) and penaltyfunction [46ndash48] can provide related optimization solutions
In this paper we proposed an improved deep-learning-hashing approach IDLH which optimized over two majorimage retrieval process
(a) In the feature extraction process the self-encodednetwork of the look-ahead type is trained by using unlabeledimage data and the expression of robust image features islearned This unlabeled learning method does not requireimage library labeling and reduces the requirements for theimage library At the same time it also takes advantage of thedeep learning networks strong learning ability and obtainsbetter image feature expression than ordinary algorithms
(b) On the index structure a secondary search is pro-posed which further increases the accuracy of the search atthe expense of very little retrieval time
Through experiments the algorithm proposed in thispaper is compared with other classic hashing algorithms onmultiple evaluation indicators Firstly we tested the learningnetworks of different code lengths and depths in order totest their effect on the retrieval system and then tested theperformance of the secondary search Through the above-mentioned series of experiments for different parameters theeffectiveness of the improved deep learning hash retrievalalgorithm proposed in this paper is verified and throughthe experimental data the good retrieval results are provedIn addition the proposed deep hashing training strategycan also be potentially applied to other hashing problemsinvolving data similarity computation
Data Availability
The data used to support the findings of this study areincluded within the article
Conflicts of Interest
The authors declare that they have no conflicts of interest
Acknowledgments
The work was funded by the National Natural ScienceFoundation of China (Grants nos 61206138 and 61373016)
References
[1] R Datta D Joshi J Li and J Z Wang ldquoImage retrieval ideasinfluences and trends of the new agerdquoACMComputing Surveysvol 40 no 2 article 5 2008
[2] G Shakhnarovich T Darrell and P Indyk Nearest-NeighborMethods in Learning andVisionTheory and PracticeMITPressCambridge MA USA 2006
12 Security and Communication Networks
[3] A Gionis P Indyk and R Motwani ldquoSimilarity search in highdimensions via hashingrdquo in 25th Int Conf pp 518ndash529 1999
[4] Z Pan J Lei Y Zhang and F L Wang ldquoAdaptive fractional-Pixel motion estimation skipped algorithm for efficient HEVCmotion estimationrdquoACMTransactions onMultimedia Comput-ing Communications and Applications (TOMM) vol 14 no 1pp 1ndash19 2018
[5] G-L Tian M Wang and L Song ldquoVariable selection in thehigh-dimensional continuous generalized linear model withcurrent status datardquo Journal of Applied Statistics vol 41 no 3pp 467ndash483 2014
[6] M Datar N Immorlica P Indyk and V S Mirrokni ldquoLocality-sensitive hashing scheme based on p-stable distributionsrdquo inProceedings of the 20th Annual Symposium on ComputationalGeometry (SCG rsquo04) pp 253ndash262 ACM June 2004
[7] B Kulis P Jain and K Grauman ldquoFast similarity search forlearned metricsrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 31 no 12 pp 2143ndash2157 2009
[8] M Raginsky and S Lazebnik ldquoLocality-sensitive binary codesfrom shift-invariant kernelsrdquo in Proceedings of the 23rd AnnualConference on Neural Information Processing Systems NIPS2009 pp 1509ndash1517 Canada December 2009
[9] L Qi X Zhang W Dou and Q Ni ldquoA distributed locality-sensitive hashing-based approach for cloud service recommen-dation from multi-source datardquo IEEE Journal on Selected Areasin Communications vol 35 no 11 pp 2616ndash2624 2017
[10] M A Carreira-Perpinan and R Raziperchikolaei ldquoHashingwith binary autoencodersrdquo in Proceedings of the IEEE Confer-ence on Computer Vision and Pattern Recognition CVPR 2015pp 557ndash566 USA June 2015
[11] J Wang S Kumar and S-F Chang ldquoSemi-supervised hashingfor large-scale searchrdquo IEEE Transactions on Pattern Analysisand Machine Intelligence vol 34 no 12 pp 2393ndash2406 2012
[12] Y Gong S Lazebnik and A Gordo ldquoIterative quantizationa Procrustean approach to learning binary codes for large-scale image retrievalrdquo in Proceedings of the IEEE Conference onComputer Vision and Pattern Recognition (CVPR rsquo11) pp 2916ndash2929 June 2011
[13] W Kong and W J Li ldquoIsotropic hashingrdquo NIPS vol 25 2012[14] M Norouzi and D J Fleet ldquoMinimal loss hashing for compact
binary codesrdquo in Proceedings of the 28th International Confer-ence on Machine Learning ICML 2011 pp 353ndash360 USA July2011
[15] J Wang W Liu A X Sun and Y-G Jiang ldquoLearning hashcodes with listwise supervisionrdquo in Proceedings of the 2013 14thIEEE International Conference on Computer Vision ICCV 2013pp 3032ndash3039 Australia December 2013
[16] G Lin C Shen Q Shi A Van Den Hengel and D Suter ldquoFastsupervised hashing with decision trees for high-dimensionaldatardquo in Proceedings of 27th IEEE Conference on ComputerVision and Pattern Recognition CVPRrsquo pp 1971ndash1978 USA2014
[17] Y Gong S Kumar H A Rowley and S Lazebnik ldquoLearningbinary codes for high-dimensional data using bilinear projec-tionsrdquo in Proceedings of the 26th IEEE Conference on ComputerVision and Pattern Recognition CVPR 2013 pp 484ndash491 USAJune 2013
[18] W Liu J Wang Y Mu and S Kumar ldquoCompact hyperplanehashing with bilinear functionsrdquo in The 29th InternationalConference on Machine Learning (ICML12) pp 467ndash474 2012
[19] Y Weiss A Torralba and R Fergus ldquoSpectral hashingrdquo inProceedings of the 22nd Annual Conference on Neural Informa-tion Processing Systems (NIPS rsquo08) pp 1753ndash1760 VancouverCanada December 2008
[20] W Liu J Wang S Kumar and S F Chang ldquoHashing withgraphsrdquo inThe 28th international conference on machine learn-ing (ICML11) 2011
[21] F Shen X Zhou Y Yang J Song H T Shen and D Tao ldquoA fastoptimization method for general binary code learningrdquo IEEETransactions on Image Processing vol 25 no 12 pp 5610ndash56212016
[22] F Shen W Liu S Zhang Y Yang and H T Shen ldquoLearningbinary codes for maximum inner product searchrdquo in Proceed-ings of the 15th IEEE International Conference on ComputerVision ICCV 2015 pp 4148ndash4156 Chile December 2015
[23] A Krizhevsky I Sutskever andG EHinton ldquoImagenet classifi-cation with deep convolutional neural networksrdquo in Proceedingsof the 26th Annual Conference on Neural Information ProcessingSystems (NIPS rsquo12) pp 1097ndash1105 Lake Tahoe Nev USADecember 2012
[24] G E Hinton and R R Salakhutdinov ldquoReducing the dimen-sionality of data with neural networksrdquo The American Associa-tion for the Advancement of Science Science vol 313 no 5786pp 504ndash507 2006
[25] A Torralba R Fergus and Y Weiss ldquoSmall codes and largeimage databases for recognitionrdquo in Proceedings of the IEEEComputer Society Conference on Computer Vision and PatternRecognition (CVPR rsquo08) pp 1ndash8 2008
[26] R Salakhutdinov andG Hinton ldquoLearning a nonlinear embed-ding by preserving class neighbourhood structurerdquo Journal ofMachine Learning Research vol 2 pp 412ndash419 2007
[27] V E Liong J Lu GWang P Moulin and J Zhou ldquoDeep hash-ing for compact binary codes learningrdquo in Proceedings of theIEEE Conference on Computer Vision and Pattern RecognitionCVPR 2015 pp 2475ndash2483 USA June 2015
[28] Y Gong S Lazebnik A Gordo and F Perronnin ldquoIterativequantization A procrustean approach to learning binary codesfor large-scale image retrievalrdquo IEEE Transactions on PatternAnalysis andMachine Intelligence vol 35 no 12 pp 2916ndash29292013
[29] W Liu J Wang R Ji Y-G Jiang and S-F Chang ldquoSupervisedhashing with kernelsrdquo in Proceedings of the IEEE Conference onComputer Vision and Pattern Recognition (CVPR rsquo12) pp 2074ndash2081 Providence RI USA June 2012
[30] J Masci A Bronstein M Bronstein and P SprechmannldquoSparse similarity-preserving hashingrdquo in Int Conf LearnRepresent pp 1ndash13 2014
[31] B Kulis and K Grauman ldquoKernelized locality-sensitive hash-ingrdquo IEEE Transactions on Pattern Analysis and Machine Intel-ligence vol 34 no 6 pp 1092ndash1104 2012
[32] F Zhao YHuang LWang and T Tan ldquoDeep semantic rankingbased hashing for multi-label image retrievalrdquo in Proceedings ofIEEE Conference on Computer Vision and Pattern RecognitionCVPR 2015 pp 1556ndash1564 June 2015
[33] G Cheng C Yang X Yao L Guo and J Han ldquoWhenDeep Learning Meets Metric Learning Remote Sensing ImageScene Classification via Learning Discriminative CNNsrdquo IEEETransactions on Geoscience and Remote Sensing pp 1ndash11
[34] J He J Feng X Liu et al ldquoMobile product search with Bag ofHash Bits and boundary rerankingrdquo in Proceedings of the 2012IEEE Conference on Computer Vision and Pattern RecognitionCVPR 2012 pp 3005ndash3012 USA June 2012
Security and Communication Networks 13
[35] F Shen Y Mu Y Yang et al ldquoClassification by retrievalBinarizing data and classifiersrdquo in Proceedings of the 40thInternational ACM SIGIR Conference on Research and Develop-ment in Information Retrieval SIGIR 2017 pp 595ndash604 JapanAugust 2017
[36] P Li S Zhao andR Zhang ldquoA cluster analysis selection strategyfor supersaturated designsrdquo Computational Statistics amp DataAnalysis vol 54 no 6 pp 1605ndash1612 2010
[37] A Pradeep S Mridula and P Mohanan ldquoHigh securityidentity tags using spiral resonatorsrdquo Cmc-Computers Materialsamp Continua vol 52 no 3 pp 187ndash196 2016
[38] Y Cao Z Zhou X Sun and C Gao ldquoCoverless informationhiding based on the molecular structure images of materialrdquoComputers Materials and Continua vol 54 no 2 pp 197ndash2072018
[39] Y LiuH Peng and JWang ldquoVerifiable diversity ranking searchover encrypted outsourced datardquo Cmc-Computers Materials ampContinua vol 55 no 1 pp 037ndash057 2018
[40] K Lin J Lu C-S Chen and J Zhou ldquoLearning compactbinary descriptors with unsupervised deep neural networksrdquo inProceedings of the 2016 IEEEConference onComputer Vision andPattern Recognition CVPR 2016 pp 1183ndash1192 USA July 2016
[41] T Do A Doan and N Cheung ldquoLearning to Hash with BinaryDeep Neural Networkrdquo in Computer Vision ndash ECCV 2016vol 9909 of Lecture Notes in Computer Science pp 219ndash234Springer International Publishing Cham 2016
[42] Rui Zhang Di Xiao and Yanting Chang ldquoA Novel ImageAuthentication with Tamper Localization and Self-Recovery inEncrypted Domain Based on Compressive Sensingrdquo Securityand Communication Networks vol 2018 Article ID 1591206 15pages 2018
[43] Xia ShuangKui and JianbinWu ldquoAModification-Free Steganog-raphy Method Based on Image Information Entropyrdquo Securityand Communication Networks vol 2018 Article ID 6256872 8pages 2018
[44] J Zhang B Qu and N Xiu ldquoSome projection-like methods forthe generalized Nash equilibriardquo Computational Optimizationand Applications vol 45 no 1 pp 89ndash109 2010
[45] Biao Qu and Jing Zhao ldquoMethods for Solving Generalized NashEquilibriumrdquo Journal of Applied Mathematics vol 2013 ArticleID 762165 6 pages 2013
[46] CWang CMa and J Zhou ldquoA new class of exact penalty func-tions and penalty algorithmsrdquo Journal of Global Optimizationvol 58 no 1 pp 51ndash73 2014
[47] Y Wang X Sun and F Meng ldquoOn the conditional andpartial trade credit policywith capital constraints A StackelbergModelrdquo Applied Mathematical Modelling vol 40 no 1 pp 1ndash182016
[48] S Lian and Y Duan ldquoSmoothing of the lower-order exactpenalty function for inequality constrained optimizationrdquo Jour-nal of Inequalities and Applications Paper No 185 12 pages2016
Chemical EngineeringInternational Journal of Antennas and
Propagation
International Journal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Navigation and Observation
International Journal of
Hindawi
wwwhindawicom Volume 2018
Advances in
Multimedia
Submit your manuscripts atwwwhindawicom
2 Security and Communication Networks
affected by dimensional disasters [3]Thehigh-level encodingcomplexity prevents widespread adoption in real-timemulti-media systems [4] One of the many practical applicationsthe approximate nearest neighbors (ANN) is very efficientHowever the method requires huge storage costs and hardto handling high-dimensional data Hence quantizationtechniques have been proposed to encode high-dimensionaldata vectors recently Due to the fact that hashing-basedANN search techniques can reduce the storage via storingthe compact binary codes hash technology is widely usedin computer vision machine learning information retrievaland other related fields in view of the retrieval of large-scaledata The goal of hash is to turn high-dimensional data intothe low-dimensional compact binary codes [5] For example119889 dimension data is turned into 119903 dimension (dgtgtr) andgenerally 119903 dimension data is between dozens of bits andhundreds of bits After turning high-dimensional data intohash code we calculate the distance or similarity betweenthe data rapidly Because of the high efficiency of binary hashcode in Hamming distance calculation and the advantages ofstorage space hash code is very efficient in large-scale imageretrieval
One of the most famous jobs in hash and the mostapplied job in the industry was locality sensitive hashing(LSH) [3] which was put forward by Gionis Indyk andMotwani in 1999[1] LSH is one of the most popular dataindependent methods which generates hash functions byrandom projections [6] In addition to traditional Euclideandistance LSH has been generalized to accommodate otherdistance and similaritymeasures such as p-normdistance [3]Mahalanobis metric [7] and kernel similarity [8] And L QiX ZhangW Dou andQ Ni put forward another applicationdirection of LSH in 2017 [9] They proposed a privacy-preserving and scalable service recommendation approachbased on distributed LSH (SerRecdistri-LSH)The purpose ofthis application is to use SerRecdistri-LSH method to handleservice recommendation in distributed cloud environment
A disadvantage of the LSH family is that LSH usuallyneeds long bit length (ge 1000) to achieve both high precisionand recallThismay leads to a huge storage overhead and thuslimits the sale at which an LSH algorithm may be appliedIn order to achieve the desired search accuracy LSH oftenrequires the long hash codes thereby reducing the recall rateThis problem can be alleviated by using multiple hash tablesbut it greatly increases storage costs and query time [10]
So learning-based data-dependent hashing methodshave become increasingly popular because of the benefitthat learned compact binary codes can effectively and highlyefficiently index and organize massive data [3] The goal ofthe data-dependent hashing methods is to generate shorthash codes (typically le 200) using available training dataVarious hashing algorithms have been proposed in theliterature of which a large category focuses on linear hashingalgorithms which learn a set of hyperplanes as linear hashfunctions The representative algorithms in this categoryinclude unsupervised PCA Hashing [11] Iterative Quanti-zation (ITQ) [12] and Isotropic Hashing [13] and super-vised Minimal Loss Hashing (MLH) [14] SemisupervisedHashing (SSH) [11] Supervised Discrete Hashing (SDH) [6]
Ranking-Based Supervised Hashing [15] FastHash [16] etcA bilinear form of hash functions was introduced by [1718]
As an extension of linear hash functions a variety ofalgorithms have been proposed to generate nonlinear hashfunctions in a kernel space including Binary ReconstructiveEmbedding (BRE) [10] RandomMaximumMargin Hashing(RMMH) [14] Kernel-Based SupervisedHashing (KSH) [12]and the kernel variant of ITQ [13] In parallel harnessingnonlinear manifold structures has been shown to be effectivein producing compact neighborhood-preserving hash codesThe early algorithm in this fashion is Spectral Hashing (SH)[19] which produces hash codes through solving a contin-uously relaxed mathematical program similar to LaplacianEigenmaps More recently Anchor Graph Hashing (AGH)[20] leveraged anchor graphs for solving the eigenfunctionsof the resulting graph Laplacians making hash code trainingand out-of-sample extension to novel data both tractableand efficient for large-scale datasets Shen et al [6 21 22]proposed a general induction
The rest of this paper is organized as follows Sec-tion 2 briefly reviews the related works about deep-learning-hashing And we present framework and steps of theimproved deep learning-based hash IDLH algorithm indetail in Section 3 Section 4 evaluates the effectiveness ofthe IDLH through a series of contrastive experiments andcarefully analyzes the experimental results followed by theconclusion in Section 5
2 Related Work
Recently deep-learning-hashing as a popular research topichas drawn increasing attention and research efforts in infor-mation retrieval computer vision and machine learningSemantic hashing is recognized as the earliest starting deeplearning hash [23]This method establishes a deep generativemodel to discover hidden binary features Such a deep modelis made as a stack of restricted Boltzmann machines (RMBs)[24] After learning a multilayer RBM by pretraining andfine-tuning the document collection the hash code of anydocument is obtained by simply thresholding the deepestoutput Such a hash code provided by deep RBM is shownto maintain a semantically similar relationship of inputdocuments into code space where each hash code (or hashkey) is used as a memory address to locate the correspondingdocument In this way semantically similar documents aremapped to adjacent memory addresses enabling efficientsearching through hash table lookups In order to improvethe performance of deep RBMS a supervised version wasproposed in [25] and the idea of nonlinear neighborhoodcomponent analysis (NCA) embedding in [26] was adoptedThe supervised information is derived from training a givenneighbornonneighbor relationship between samples Thenbased on the depth RBM the objective function of the NCAis optimized so that the depth RBM yields a hash codeNote that supervised deep RBMS can be applied to widedata fields other than text data In [25] the depth RBMs aresupervised using a Gaussian distribution and the models of
Security and Communication Networks 3
the visible units in the first layer are successfully applied tothe processing of massive image data
In [27] a deep neural network was developed to learnmultilevel nonlinear transformations mapping the originalimage to a compact binary hash code to support large-scaleimage retrieval for learning binary image representationsA deep hashing model is established under the three-layerconstraint on the deep network (1) the reconstruction errorbetween the original real-valued image feature vector andthe resulting binary code is minimized (2) each bit of thebinary code has a balance (3) all bits are independent of eachother Similar constraints were used in previous unsupervisedhash or binary codingmethods such as iterative quantization(ITQ) [28] In [27] a supervised version is called superviseddepth hash 3 in which a discriminative item containing twopairs of supervised information is added to the objectivefunction of the deep hash model
It is worth noting that in the training of sparse neuralnetworks in addition to the sparse similarity maintaininghash depth hash and supervised depth hash the pretrainingphase is not included Instead the hash codes are learnedfrom scratch using a set of training data However nopretraining can make the general hash code less efficient Inparticular the sparse similarity keep hash method is foundto be inferior to the existing supervised hashing methodie kernel-based supervised hashing (KSH) [29] in terms ofsearch accuracy on some image datasets [30] the deep hashmethod and its supervised version are slightly better than ITQand its supervised version CCA+ ITQ respectively [31] Notethat KSH ITQ and CCA+ITQ develop a shallow learningframework
One of the main purposes of deep learning is to learnthe robust and powerful representation of complex data Itis very natural to use deep learning to explore compact hashcodes which can be thought of as binary representation ofthe data The deployed CNN consists of three convolutionpools including rectified linear activation maximum poolmerging and local contrast normalization a standard fullyconnected layer and an output layer with softmax activationfunctions In [32] a newmethod called deep semantic sortinghashing is proposed to learn hash values therebymaintainingmultilevel semantic similarity between multilabel imagesThis method is combined with convolutional neural networkhashing method taking image pixels as input training depthCNN and jointly learning image feature representation andhash value by thismethodThe deployed CNN consists of fiveconvolution-pooling layers two fully connected layers and ahash layer (ie output layer)
Indexing massive amounts of multimedia data such asimages and videos is a natural application based on learninghashing In particular due to the well-known semanticdivide hashing methods have been extensively studied forimage search and retrieval as well as mobile product search[33 34] Although hashing techniques have been appliedto the active learning framework to cope with big dataapplications these hash algorithms are rarely applied toimage classification For the image recognition problem withmany categories the computational and memory overheadprimarily stems from the large number of classifiers to be
Table 1 A huge number of parameters need to be learned andstored with different datasets Considering a classification task withC different categories and D-dimensional feature representationeven the simplest linear models are comprised of DtimesC parameters
learned Table 1 show that the complexities can be high atthe stages of both training and deploying these classifiersInspired by [35 36] we proposed a combination classifierswith DNN and binarizing classifiers to solve some classicsecurity calculation problems [37ndash39] Different from otherdeep learning-hash methods that often entails heavy compu-tations by using a conventional classifier exemplified by K-NNand SVM ourmethod learns classifiers using binary hashcodes which are simultaneously learned from the trainingdata The combination classifiers can provide both imagefeatures and accelerate image classification and thus it makethe large-scale image recognition faster The advantages ofthe extending hashing techniques from fast image search toimage classification inspire us to apply them to deep learninghash framework
With the development of Internet technology the spread-ing form of illegal information on the Internet is changinggradually Traditionally the form of communication basedon word description has been transformed into a diversifiedformof communication based on video and imageThereforethe original keyword blocking web content grading blacklistrestriction and other filtering methods have not been able toblock illegal information dissemination At present networkillegal image recognition is mainly divided into the followingcategories
(1) Erotic Image Recognition The identification of onlinepornographic imagesmainly includesmethods based on skincolor feature extraction detection and judgment andmethodsbased on limb judgment The limb state is determinedby extracting the divided skin color information and theconnection feature of the human body posture and furtherdetermining whether the image transmitted on the networkcontains pornographic content
(2) Images Involving State Secrets or Military Secrets Therecognition of secret-related images mainly involvessteganography which hides secret information in imagevideo and other carriers for secret transmission At presentthe high-order statistical features based on modeling thecomplex correlation of image neighborhood become themainstream features in the field of steganalysis SRM(Spatial Rich Model) PSRM (Projection Speciation RichModel) and other models are based on such high-orderhigh-dimensional features and have achieved good detectionresults Steganalysis based on depth learning is a hotspot in
4 Security and Communication Networks
the field of information hiding in order to identify the illegalsecret-related images accurately and quickly
(3) Images Containing Antihuman Content such as TerroristViolenceThe identification of such images is mainly based onimage contrast techniques Image comparison includes tech-niques such as image feature extraction high-dimensionalspatial feature index establishment and similarity measureIt is a very worthwhile to study how to quickly comparethe massive network images to the illegal target images sothat the recall rate and the precision rate can be taken intoaccount
3 The Proposed Method
In this section we will present the notations as summarizedin Table 2 firstly The concept of deep learning stems fromthe field of artificial neural networks Deep learning is deepneural network learning and is a learning structure withmultiple hidden layers In the process of deep learningthe network is trained layer by layer Each layer of thelearning network extracts certain features and informationand takes the training result as a deeper input Finallythe entire network is fine-tuned with a top-down algo-rithm
Through deep learning complex function expressionscan be learned thereby completing the concept of high-levelabstraction from the underlying information It has beenwidely used in language understanding target recognitionand speech perception
Figure 1 shows that the proposed framework IDLHincludes three components The first component is prepro-cessing layer on the image dataset The second component isthe training self-coding network with a layer-by-layer greedylearning algorithm to obtain the feature expression functionof the image The third is hash layer which retrieves imagessimilar to the query image with compact binary codes andcategorizes the query one by the majority semantic labelwithin the hashing bucket
31 Preprocessing Since the depth learning algorithm usedin this paper is an unsupervised learning algorithm it canautomatically learn the deep features of the image fromthe original pixel information of the image Therefore theoriginal pixel value of the whole image can be directly used asinput data for the deep learning model In order to facilitatethe training of the network it is necessary to preprocess theimage Through preprocessing the image is simply scaledsample-by-sample mean-value reduction and whitening isprocessed to reduce the redundant information in the imageand facilitate the deep learning network for training andcalculation Preprocessing can be further grouped into threesuboperations
(1) Normalization Normalization can prevent neuron outputsaturation caused by excessive net input absolute value Weuse the sigmoid function to do normalization as shown asfollows
Table 2 Summary of notations
Symbol Definition119909119894 the gray value of the image pixels120583(119894) mean pixels
U an arbitrary orthogonal matrix and defines in theZCA whitening
J(Wb) the quantization loss between the learned binaryvalues and the real values
120582 a weight attenuation parameter119878119894 119894119905ℎ seta b two thresholdsD dimensionality of data pointsM K number of candidate imagesE the objective function
119909119894 = 1(1 + 119890minus119909119894) (1)
119909119894 is the gray value of the image pixels
(2) Pointwise Mean Reduction This process mainly is to getrid of the redundant information of the image the meanvalue is eliminated for each point of the image the averagebrightness of the image is removed and the DC componentof the data is eliminated Assuming that 119909(119894) isin 119877119899 is the grayvalue of each pixel of image I we use formulae (2) and (3) tozero-mean image
120583(119894) = 1119899119899sum119895=1
119909(119894)119895 (2)
119909(119894)119895 = 119909(119894)119895 minus 120583(119894) (3)
(3) Whitening Whitening is an important pretreatmentprocess its purpose is to reduce the redundancy of input dataso that the whitened input data has the following properties(i) low correlation between features (ii) all features having thesame variance then the formula is as formula (4)
In formula (4) the rotation matrix of 119909119903119900119905119894 is 119880119879119909119894Generally when 119909 is in interval [-11] 120576 asymp 10minus5
32 Training Stack Sparse Self-Encoding Network Stack self-coding neural network has strong expressive ability whichmainly benefits from its hierarchical feature representationThrough one level of feature learning we can learn thehierarchical structure between features Stack self-encodingneural network is a neural network model composed ofmultilayer sparse self-encoder that is the output of theformer self-encoder as the input of the latter self-encoder
In the training the original input 119909(119896) is used as inputto train the first self-encoded neural network At this pointfor each training sample 119909(119896) the output ℎ(119896)1 of the hidden
Security and Communication Networks 5
Hash Layer 2
Hamming Retrieval
Query Result
Weights
Codes
Training Dataset
Feature expression
function
Image Preprocess Layer Feature
expression function
Hash Layer 1
Binary Weights
Binary Codes
Feature migration
Query Image
Figure 1 Deep learning-hash retrieval framework IDLH consists of three main components (preprocessing layer deep neural networklayer and hash layer) The object of the first layer is simply scaled sample-by-sample mean-value reduction and whitening In the secondcomponent we develop a deep neural network to obtain the feature expression function of the image And the classifier weights and featurebinary codes are simultaneously learned in the last component-hash layer
layer can be obtained and the output of the hidden layer canbe used as the input of the second self-encoder to continuetraining the second self-encoder Then the output ℎ(119896)2 of thesecond hidden layer of the self-encoder can be obtained Theoutputℎ(119896)1 of the first hidden layer of the self-encoder is calleda first-order feature and the output ℎ(119896)2 of the second hiddenlayer of the self-encoder is called a second-order feature Inorder to classify the two-order feature ℎ(119896)2 can be used as theinput of Softmax regression
Figure 2 shows the flowchart of the proposed methodAnd there are mainly three processes (supervised pretrain-ing rough image retrieval and accurate image retrieval)The object of the first process is to transform the high-dimensional feature vector into a low-dimensional compacttwo value codes through hash function In the secondprocedure we pick out M candidate images by calculatingHamming distance In the third process we calculate theEuclidean distance between the candidate image and theimage to be retrieved and accurately extract K images fromthe M candidate images
Figure 3 shows a block diagram of a self-encoding neuralnetwork The stacking self-encoding network contains 3hidden layers (feature layers) The input layer inputs theoriginal data i into the first layer of the feature layer theoutput result of the former layer serves as the input of thenext layer and the output of the third layer serves as thefeature expression of the image In our method it is also usedas the input of the hash classifier and it is possible to usethe characteristics of the STD neural network to classify thefeatures
By using the matrix representation of the binary codesvectors and the output of the 3th layer of the network we usethe gradient descent method to solve the neural network
For a single sample (x y) the cost function is as shown in
(12 10038171003817100381710038171003817ℎ119882119887 (119909(119894)) minus 119910(119894)100381710038171003817100381710038172)]
+ 1205822119899119897minus1sum119897=1
119904119897sum119894=1
119904119897+1sum119895=1
(119882(119897)119895119894 )2
(6)
The first term 119869(119882 119887) represents the mean variance termThe second term aims to prevent the data from overfitting byreducing the magnitude of the weight 120582 is a weight attenua-tion parameter It is used to balance the relative importance ofmean square deviation terms and weight attenuation termsOur purpose is to minimize the quantization loss 119869(119882 119887)between the learned binary values and the real values of aninput image according to parameters119882 and 11988733 Hash Algorithm Retrieval The image retrieval methodbased on hash algorithmmaps the high-dimensional contentfeatures of images into Hamming space (binary space) andgenerates a low-dimensional hash sequence to represent apicture This method reduces the requirement of computermemory space for image retrieval system improves theretrieval speed and better adapts to the requirements of massimage retrieval
Inspired by [6 8] we use a set of hash functions to hashdata into different buckets After we do some hash mappingon the original image feature data we hope that the originaltwo adjacent feature data can be hash into the same bucketwith the same bucket number And then after hash mappingof all the data in the original feature set we can get a hashtable These original feature data sets are scattered into hashtable buckets and the data belonging to the same bucket isprobably adjacent to the original data However there is also
6 Security and Communication Networks
StartTwo valued by the
Sigmod function in Hash layer
Building hash pool and hash bucket
Calculation of Hamming distance
Select the top m candidate images
Calculation of Euclidean distance
Select the top kimages from m
candidate imagesas retrieval result
End
Supervised pre-training
Rough image retrieval
If itgt05Yes
1
No
0
If it less than a threshold
Yes
No
Accurate image retrieval
The featureof the image
to be retrieved
Self-learning network
The featureof the
training images
Binary processing
Binary codes for trainingimage
Binary codes for retrieving
image
Figure 2 Deep learning-hash retrieval flowchart IDLHmainly includes three processes (supervised pretraining rough image retrieval andaccurate image retrieval)
+1+1 +1
+1
Layer 0Input layer
Layer 1 Layer 2 Layer 3 Hash Layer
x1
x2
x3
x4
x5
ℎ(1)1
ℎ(1)2
ℎ(1)3
ℎ(1)4
ℎ(2)1
ℎ(2)2
ℎ(2)3
ℎ(2)4
ℎ(3)1
ℎ(3)2
ℎ(3)3
Output ℎ(3)k
from layer 3 is usedas binary code with bk = f(ℎ(3)
k)
Figure 3 Self-learning network based on stack self-encoding network The neurons labeled xi is the input of the neural network and ldquo+ 1rdquoare the offset nodes (intercept entries) of the neural network The layer 0 is the input layer of neural network and layer 3 is the output layerof neural network The middle layers of layer 0 layer to layer 3 are the hidden layers of neural network
a small probability in events that is the nonadjacent data ishash to the same barrel Set the hash function as the follows
Here119908119896 is the projection vector and 119887119896 is the correspond-ing intercept The code value generated by formula (7) isminus1 1 and we use the following formula to convert it intotwo value codes
119910119896 = 12 (1 + ℎ119896 (119909)) (8)
Given a sample point isin 119877119863 we can compute a K-bitbinary code 119910 for 119909 with formula (9) The hash functionperforms the mapping as ℎ119896 119877119863 997888rarr 119861
Here 119883 = 119909119899119873119899=1 isin 119877119863times119873 is the feature data matrix withpoints as columns Such a binary encoding process can also beviewed as mapping the original data point to a binary valuedspace namely Hamming space
34 SimilarityMeasure After obtaining the binary hash codeof the image it is necessary tomeasure similarity between theretrieved image and the library image in the Hamming spaceThe smaller the Hamming distance is the closer distancebetween the two data is and the degree of similarity is higherotherwise the two data similarity is lower
Here oplus is an XOR operation The two sets of 119910119894 and 119910119895represent the hash code of the search image feature and theimage library is mapped through the hash function The newimage features learned by the stack self-encoding network aregenerated by the hash function The storage structure of theimage feature vectors is shown in Table 3
As can be seen from Table 3 the hash code of the imageis related to the image ID and the image name one by one Inthe process of searching the image feature vector is obtainedthrough deep learning by a hash function the original data ismapped into a newdata space and a corresponding hash codeis obtained The hash code is used to calculate the Hammingdistance in the Hamming space as a measure of similaritybetween images Finally the storage structure of the imagefeature vector is used to find the corresponding image ID ofthe hash code and the output retrieval result is output to theuser
35 Image Secondary Search In the first-level search phasethe features learned from the deep learning network aremapped into the Hamming space using the hash function Inthe similarity measurement phase the traditional Euclideandistance is abandoned Measure the similarity betweenimages by comparing the Hamming distance between theimage features of the query image and the image of thelibrary image In order to further improve the accuracy ofretrieval without affecting the real-time performance we canretrieve the image by the second level retrieval These stepsare described in detail as follows After one level retrievalwe choose the K images with the most similarity in thefirst-level retrieval result and then calculate the Euclideandistance between the original feature vector of the K imagesand the original feature vector of the query imageThe resultsobtained as the similarity measure of the images and outputthe retrieval result that has been ranked from the high andlow with the similarity distance
Although theHash algorithmmaps the high-dimensionalfeature vectors of the image into a hash-coded form theproblem of ldquodimensional disastersrdquo is solved and the retrievalefficiency is greatly accelerated However when the similaritycomparison is performed the Hamming distances of theimage features are simply compared using the results of theprimary search and occasionally undesirable results maystill appear on the search results If we want to increasethe accuracy of the search we must increase the hash codelength However excessively long codes will increase theamount of calculations increase the memory burden andreduce the real-time nature of retrieval failing to achieve
the goal of reducing the size of data In order to solve thisproblem keep the retrieval efficiency and further improvethe retrieval accuracy we propose a search strategy forsecondary retrieval the specific steps of which are as follows
Step 1 Through the first-level search in the Hamming spacethe similarity degree of the images is sorted and the top Ksorting images are selected
Step 2 For the 119896 images in Step 1 calculate the Euclideandistance one by one from its original image feature vector tothe image feature vector of the query image
Step 3 The Euclidean distance calculated in Step 2 is sortedThe smaller the calculated value is the higher similaritybetween images is and the similarity is sorted from high tolow and output as the final search result
In the second search it is necessary to pay attention tothe selection of the 119870 value although the larger the 119870 valueis the better the search effect is but accordingly the longerthe time is consumed Therefore it is necessary to combinevarious factors to select the appropriate119870 value
4 Experimental and Performance Analysis
In this section we thoroughly compare the proposedapproach with the improved deep learning hash retrievalmethods on several benchmark datasets Through a series ofexperiments the effectiveness and feasibility of the proposedalgorithm are verified
41 Database Two mostly used databases in the recent deeplearning hash works are taken into our evaluation Thetwo image libraries are derived from the CIFAR-10[11] coreexperimental image library dataset and theCaltech 256 imagelibrary dataset
CIFAR-10 dataset contains 10 object categories and eachclass consists of 6000 images resulting in a total of 60000images The dataset is split into training and test sets whichare averagely divided into 10 object classes The Caltech 256image library dataset contains 29780 color images which aregrouped intro 256 classes
First test the CIFAR-10 image dataset There are a totalof 50000 training samples which are used for training on thedeep learning networkThe remaining 10000 images are usedas test samples And then we randomly select 50 images fromdatabase as the query images For theHidden Image Retrievalalgorithm based on deep learning mentioned in this paperthe image pixel data is directly used as input while for otheralgorithms the 512-dimensional GIST feature is used as thefeature expression of the image Note quantization all imagesinto 32lowast32 sizes before experiment
For the Caltech 256 image a total of 256 classes areincluded and each class contains at least 70 images There-fore 70 images of each class a total of 17920 images arerandomly selected and are used as training images Theremaining images are used as test samples In addition all ofthe imagesrsquo size is set to 64lowast64 again when training
8 Security and Communication Networks
42 Evaluation Metrics We measure the performance ofcompared methods using Precision-recall and Average-Retrieval Precision (ARP) curves Precision is the ratio ofthe correct number of images m in the search result to thenumber k of all returned images The formula is as follows
precision = 119898119896 times 100 (12)
Recall is the ratio of the correct number of imagesm in thesearch results to the number g of images in the image libraryThe formula is as follows
recall = 119898119902 times 100 (13)
Assume that the search result of the query image 119894 is119861119894 and 119860 119894 means that the category is the same between thequery image and the return image then the accuracy rate forthe image query result 119875(119894) can be defined by the followingformula
119875 (119894) = |119860 (119894) cap 119861 (119894)||119861 (119894)| (14)
Average-Retrieval Precision (ARP) the average value ofall the images in the same class as the retrieval rate obtainedfrom the retrieval image is defined as follows
Here 119868119863119898 is the category index number of the image119898 isthe category index119873 is the number of imageswhose categoryis 119868119863119898 and 119894119889(119894) is the category index number of the queryimage
43 Performance Analysis In the proposed algorithm IDLHthe length of the hash sequence and the depth of the hiddenlayer in the deep learning network are two key parametersWhen the hash sequence length is small different featurevectors can easily be mapped into the same hash sequence sothe retrieval accuracy is low However if the hash sequence istoo long a large storage space is required and a long time isconsumed which reduces the real-time performance For thenumber of hidden layers the number of layers in the hiddenlayer is too small which is not conducive to learning strongimage features However if the depth of the hidden layer istoo large the difficulty of training is increased In order toverify the effectiveness and feasibility of our algorithm weconducted the following experiments
(1) Results on CIFAR-10 Dataset Figure 4(a) shows the searchresults of the Average-Retrieval Precision using our proposedalgorithm IDLH compared with the LSH algorithm [3] andother three deep learning algorithms the DH algorithm [21]the DeepBit algorithm [40] and the UH-BDNN algorithm[41] on CIFAR-10 dataset with 8 16 32 48 64 96 128and 256 bits Figure 4(b) shows the Precision-recall curveunder 48-bit encoding It can be seen that the algorithm hasa higher precision than the other hashing algorithms with
the same recall rate However the advantage is not obviousand the average accuracy is slightly higher than other hashalgorithms
In order to overcome the above defects we use deeplearning to perform hash mapping on image features per-form hash encoding of different bits on the same featurecalculate the Precision-recall of the search results under thecondition of different coded bits and determine the impactof the encoding length on the retrieval results
As Figure 5 shows with the increase in the number ofcoded bits the Precision-recall is continuously increasingWith the increase in the number of coded bits the imageis better expressed However after the number of coded bitsreaches 64 even if the number of coded bits increases theaverage accuracy rate increases relatively slowly Because theinformation of the tiny image is relatively simple when thenumber of encoded bits reaches 64 bits a relatively goodimage expression has been obtained and the performanceof the algorithm has basically stabilized At this time despiteincrease in the number of encoding bits it is not very helpfulto improve the accuracy rate
In addition we want to test the influence of the numberof hidden layers in the deep learning network on the retrievalresult by changing the number of hidden layers
Figure 6 shows the effect of deep learning networks onexperimental results in the case of different hidden layernumbers
It can be seen that deeper networks do not have muchimprovement in performance which is different from theexpectation that more hidden layers will help learn strongerimage features Since the image library data used in theexperiment is a tiny image library relatively good image char-acteristics can be learned using a deep learning network withfewer layers However if the image library is replaced witha more colorful image the deep neural network can acquiremore detailed image features and the deepening of thelearning network will greatly help the study of image features
(2) Results on Caltech 256 Image Data Set Figure 7(a)shows the results of the Average-Retrieval Precision resultswhen the number of coded bits is different Compared withthe black and white image library the proposed algorithmembodies the advantage of image feature learning and leadsthe Average-Retrieval Precision to other hash retrieval algo-rithms In Figure 7(b) we can also see that the algorithmproposed in this paper has a higher Precision-recall thanother algorithms under the same recall rate and it has bettersearch performance
As shown in Figure 8 as the number of coded bitsincreases the precision rate increases with the same recallrate This feature of the color image library is more pro-nounced than the black and white image library Becausethe color image contains more information more codingis needed to express it and the increase of encoding helpsto learn the features of the image The experimental resultsalso show that the deep learning network has learned moreexcellent image features
In Figure 9 the precision rate is significantly improvedby the increase in the number of hidden layers in the
Security and Communication Networks 9
IDLHLSHDHDeepBitUH-BDNN
16 32 48 64 96 128 2568Bits
0
01
02
03
04
05
06
07
08Av
erag
e Ret
rieva
l Pre
cisio
n
(a) Average-Retrieval Precision
IDLHLSHDHDeepBitUH-BDNN
01 02 03 04 05 06 07 08 09 100Recall
0
01
02
03
04
05
06
07
08
09
Prec
ision
(b) Precision-recall at 48 bits
Figure 4 Five kinds of algorithm retrieval performance comparison on CIFAR-10 dataset
8 bits16 bits32 bits48 bits64 bits128 bits
01 02 03 04 05 06 07 08 09 10Recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
Figure 5 Precision-recall curves with lengths
color image library Caltech 256 This is because theinformation contained in a more colorful image is morecomplex Adding a hidden layer can learn more detailsof the image and help improve the accuracy of thesearch
Next we tested the performance of secondary imageretrieval The value of k in the secondary search is 20 andthe experimental results are shown in Figure 10 As can
Level 1Level 2Level 3Level 4
01 02 03 04 05 06 07 08 09 1000Recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
Figure 6 Precision-recall curves with different code differenthidden layers
be seen from the results secondary retrieval can effectivelyimprove the retrieval accuracy when the number of codedbits is small However with the increase in the number ofencoding bits the results of the secondary search and theaccuracy of the primary search are not much different Thisis because the shorter the hash sequence is the easier thefeature vectors with different original features are mappedto the same hash code In order to make up for the errors
10 Security and Communication Networks
IDLHLSHDHDeepBitUH-BDNN
16 32 48 64 96 128 2568Bits
0
01
02
03
04
05
06
07
08
09
1Av
erag
e-Re
trie
val P
reci
sion
(a) Average-Retrieval Precision
IDLHLSHDHDeepBitUH-BDNN
01 02 03 04 05 06 07 08 09 1000Precision-recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
(b) Precision-recall at 48 bits
Figure 7 Five kinds of algorithm retrieval performance comparison on Caltech 256 set
8 bits16 bits32 bits48 bits64 bits128 bits
01 02 03 04 05 06 07 08 09 10Recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
Figure 8 Precision-recall curves with different code lengths
caused by the short hash code it is necessary to performsecondary searchWhen the number of encoding bits is smalla secondary retrieval method is used in IDLH and the searchaccuracy rate can be improved at the expense of a small searchspeed
Level 1Level 2Level 3Level 4
01 02 03 04 05 06 07 08 09 10Recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
Figure 9 Precision-recall curves with different hidden layers
5 Conclusion
With the rapid development of data storage and digital pro-cess more and more digital information is transformed andtransmitted over the Internet day by day which brings peoplea series of security problems as well as convenience [42] Theresearches on digital image security that is image encryptionimage data hiding and image authentication become moreimportant than ever [38 39] The most essential problem of
Security and Communication Networks 11
First retrievalSecond retrieval
16 32 48 64 96 128 2568Bits
07
08
09
1
Aver
age-
Retr
ieva
l Pre
cisio
n
Figure 10 Average-Retrieval Precision with first and secondretrieval
image recognition is to extract robust features The qualityof feature extraction is directly related to the effect of recog-nition so most of the previous work on image recognitionis spent on artificial design features [43] In recent yearsthe emergence of deep learning technology has changed thestatus of artificial design classification characteristics Deeplearning technology simulates the mechanism of humanvisual system information classification processing from themost primitive image pixels to lower edge features then tothe target components that are combined on the edge andfinally to the whole target depth learning can be combinedby layer by layer The high-level feature is the combination oflow level features From low level to high level features aremore and more abstract and show semantics more and moreFrom the underlying features to the combination of high-levelfeatures it is the depth of learning that is done by itself Itdoes not require manual intervention Compared with thecharacteristics of the artificial design this combination offeatures can be closer to the semantic expression
In terms of illegal image retrieval the traditional recog-nition method should establish a recognition model foreach type of recognition task In the actual application arecognition model needs a recognition server If there aremany identification tasks the cost is too high We used thedeep neural network to recognize the illegal image it onlyneeds to collect the samples of every kind of illegal imageand participate in the training of the deep neural networkFinally a multiclassification recognition model is trainedWhen classifying unknown samples deep neural networkaccounting calculates the probability that the image belongsto each class
We all know that in the image detection process theaccuracy and recall rate are mutually influential Ideally bothmust be high but in general the accuracy is high and the
recall rate is low the recall rate is high and the accuracy islow For image retrieval we need to improve the accuracyunder the condition of guaranteeing the recall rate Forimage disease surveillance and anti-illegal images we need toenhance the recall under the condition of ensuring accuracyTherefore in different application scenarios in order toachieve a balance between accuracy and recall perhaps somegame theory (such as Nash Equilibrium [44 45]) and penaltyfunction [46ndash48] can provide related optimization solutions
In this paper we proposed an improved deep-learning-hashing approach IDLH which optimized over two majorimage retrieval process
(a) In the feature extraction process the self-encodednetwork of the look-ahead type is trained by using unlabeledimage data and the expression of robust image features islearned This unlabeled learning method does not requireimage library labeling and reduces the requirements for theimage library At the same time it also takes advantage of thedeep learning networks strong learning ability and obtainsbetter image feature expression than ordinary algorithms
(b) On the index structure a secondary search is pro-posed which further increases the accuracy of the search atthe expense of very little retrieval time
Through experiments the algorithm proposed in thispaper is compared with other classic hashing algorithms onmultiple evaluation indicators Firstly we tested the learningnetworks of different code lengths and depths in order totest their effect on the retrieval system and then tested theperformance of the secondary search Through the above-mentioned series of experiments for different parameters theeffectiveness of the improved deep learning hash retrievalalgorithm proposed in this paper is verified and throughthe experimental data the good retrieval results are provedIn addition the proposed deep hashing training strategycan also be potentially applied to other hashing problemsinvolving data similarity computation
Data Availability
The data used to support the findings of this study areincluded within the article
Conflicts of Interest
The authors declare that they have no conflicts of interest
Acknowledgments
The work was funded by the National Natural ScienceFoundation of China (Grants nos 61206138 and 61373016)
References
[1] R Datta D Joshi J Li and J Z Wang ldquoImage retrieval ideasinfluences and trends of the new agerdquoACMComputing Surveysvol 40 no 2 article 5 2008
[2] G Shakhnarovich T Darrell and P Indyk Nearest-NeighborMethods in Learning andVisionTheory and PracticeMITPressCambridge MA USA 2006
12 Security and Communication Networks
[3] A Gionis P Indyk and R Motwani ldquoSimilarity search in highdimensions via hashingrdquo in 25th Int Conf pp 518ndash529 1999
[4] Z Pan J Lei Y Zhang and F L Wang ldquoAdaptive fractional-Pixel motion estimation skipped algorithm for efficient HEVCmotion estimationrdquoACMTransactions onMultimedia Comput-ing Communications and Applications (TOMM) vol 14 no 1pp 1ndash19 2018
[5] G-L Tian M Wang and L Song ldquoVariable selection in thehigh-dimensional continuous generalized linear model withcurrent status datardquo Journal of Applied Statistics vol 41 no 3pp 467ndash483 2014
[6] M Datar N Immorlica P Indyk and V S Mirrokni ldquoLocality-sensitive hashing scheme based on p-stable distributionsrdquo inProceedings of the 20th Annual Symposium on ComputationalGeometry (SCG rsquo04) pp 253ndash262 ACM June 2004
[7] B Kulis P Jain and K Grauman ldquoFast similarity search forlearned metricsrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 31 no 12 pp 2143ndash2157 2009
[8] M Raginsky and S Lazebnik ldquoLocality-sensitive binary codesfrom shift-invariant kernelsrdquo in Proceedings of the 23rd AnnualConference on Neural Information Processing Systems NIPS2009 pp 1509ndash1517 Canada December 2009
[9] L Qi X Zhang W Dou and Q Ni ldquoA distributed locality-sensitive hashing-based approach for cloud service recommen-dation from multi-source datardquo IEEE Journal on Selected Areasin Communications vol 35 no 11 pp 2616ndash2624 2017
[10] M A Carreira-Perpinan and R Raziperchikolaei ldquoHashingwith binary autoencodersrdquo in Proceedings of the IEEE Confer-ence on Computer Vision and Pattern Recognition CVPR 2015pp 557ndash566 USA June 2015
[11] J Wang S Kumar and S-F Chang ldquoSemi-supervised hashingfor large-scale searchrdquo IEEE Transactions on Pattern Analysisand Machine Intelligence vol 34 no 12 pp 2393ndash2406 2012
[12] Y Gong S Lazebnik and A Gordo ldquoIterative quantizationa Procrustean approach to learning binary codes for large-scale image retrievalrdquo in Proceedings of the IEEE Conference onComputer Vision and Pattern Recognition (CVPR rsquo11) pp 2916ndash2929 June 2011
[13] W Kong and W J Li ldquoIsotropic hashingrdquo NIPS vol 25 2012[14] M Norouzi and D J Fleet ldquoMinimal loss hashing for compact
binary codesrdquo in Proceedings of the 28th International Confer-ence on Machine Learning ICML 2011 pp 353ndash360 USA July2011
[15] J Wang W Liu A X Sun and Y-G Jiang ldquoLearning hashcodes with listwise supervisionrdquo in Proceedings of the 2013 14thIEEE International Conference on Computer Vision ICCV 2013pp 3032ndash3039 Australia December 2013
[16] G Lin C Shen Q Shi A Van Den Hengel and D Suter ldquoFastsupervised hashing with decision trees for high-dimensionaldatardquo in Proceedings of 27th IEEE Conference on ComputerVision and Pattern Recognition CVPRrsquo pp 1971ndash1978 USA2014
[17] Y Gong S Kumar H A Rowley and S Lazebnik ldquoLearningbinary codes for high-dimensional data using bilinear projec-tionsrdquo in Proceedings of the 26th IEEE Conference on ComputerVision and Pattern Recognition CVPR 2013 pp 484ndash491 USAJune 2013
[18] W Liu J Wang Y Mu and S Kumar ldquoCompact hyperplanehashing with bilinear functionsrdquo in The 29th InternationalConference on Machine Learning (ICML12) pp 467ndash474 2012
[19] Y Weiss A Torralba and R Fergus ldquoSpectral hashingrdquo inProceedings of the 22nd Annual Conference on Neural Informa-tion Processing Systems (NIPS rsquo08) pp 1753ndash1760 VancouverCanada December 2008
[20] W Liu J Wang S Kumar and S F Chang ldquoHashing withgraphsrdquo inThe 28th international conference on machine learn-ing (ICML11) 2011
[21] F Shen X Zhou Y Yang J Song H T Shen and D Tao ldquoA fastoptimization method for general binary code learningrdquo IEEETransactions on Image Processing vol 25 no 12 pp 5610ndash56212016
[22] F Shen W Liu S Zhang Y Yang and H T Shen ldquoLearningbinary codes for maximum inner product searchrdquo in Proceed-ings of the 15th IEEE International Conference on ComputerVision ICCV 2015 pp 4148ndash4156 Chile December 2015
[23] A Krizhevsky I Sutskever andG EHinton ldquoImagenet classifi-cation with deep convolutional neural networksrdquo in Proceedingsof the 26th Annual Conference on Neural Information ProcessingSystems (NIPS rsquo12) pp 1097ndash1105 Lake Tahoe Nev USADecember 2012
[24] G E Hinton and R R Salakhutdinov ldquoReducing the dimen-sionality of data with neural networksrdquo The American Associa-tion for the Advancement of Science Science vol 313 no 5786pp 504ndash507 2006
[25] A Torralba R Fergus and Y Weiss ldquoSmall codes and largeimage databases for recognitionrdquo in Proceedings of the IEEEComputer Society Conference on Computer Vision and PatternRecognition (CVPR rsquo08) pp 1ndash8 2008
[26] R Salakhutdinov andG Hinton ldquoLearning a nonlinear embed-ding by preserving class neighbourhood structurerdquo Journal ofMachine Learning Research vol 2 pp 412ndash419 2007
[27] V E Liong J Lu GWang P Moulin and J Zhou ldquoDeep hash-ing for compact binary codes learningrdquo in Proceedings of theIEEE Conference on Computer Vision and Pattern RecognitionCVPR 2015 pp 2475ndash2483 USA June 2015
[28] Y Gong S Lazebnik A Gordo and F Perronnin ldquoIterativequantization A procrustean approach to learning binary codesfor large-scale image retrievalrdquo IEEE Transactions on PatternAnalysis andMachine Intelligence vol 35 no 12 pp 2916ndash29292013
[29] W Liu J Wang R Ji Y-G Jiang and S-F Chang ldquoSupervisedhashing with kernelsrdquo in Proceedings of the IEEE Conference onComputer Vision and Pattern Recognition (CVPR rsquo12) pp 2074ndash2081 Providence RI USA June 2012
[30] J Masci A Bronstein M Bronstein and P SprechmannldquoSparse similarity-preserving hashingrdquo in Int Conf LearnRepresent pp 1ndash13 2014
[31] B Kulis and K Grauman ldquoKernelized locality-sensitive hash-ingrdquo IEEE Transactions on Pattern Analysis and Machine Intel-ligence vol 34 no 6 pp 1092ndash1104 2012
[32] F Zhao YHuang LWang and T Tan ldquoDeep semantic rankingbased hashing for multi-label image retrievalrdquo in Proceedings ofIEEE Conference on Computer Vision and Pattern RecognitionCVPR 2015 pp 1556ndash1564 June 2015
[33] G Cheng C Yang X Yao L Guo and J Han ldquoWhenDeep Learning Meets Metric Learning Remote Sensing ImageScene Classification via Learning Discriminative CNNsrdquo IEEETransactions on Geoscience and Remote Sensing pp 1ndash11
[34] J He J Feng X Liu et al ldquoMobile product search with Bag ofHash Bits and boundary rerankingrdquo in Proceedings of the 2012IEEE Conference on Computer Vision and Pattern RecognitionCVPR 2012 pp 3005ndash3012 USA June 2012
Security and Communication Networks 13
[35] F Shen Y Mu Y Yang et al ldquoClassification by retrievalBinarizing data and classifiersrdquo in Proceedings of the 40thInternational ACM SIGIR Conference on Research and Develop-ment in Information Retrieval SIGIR 2017 pp 595ndash604 JapanAugust 2017
[36] P Li S Zhao andR Zhang ldquoA cluster analysis selection strategyfor supersaturated designsrdquo Computational Statistics amp DataAnalysis vol 54 no 6 pp 1605ndash1612 2010
[37] A Pradeep S Mridula and P Mohanan ldquoHigh securityidentity tags using spiral resonatorsrdquo Cmc-Computers Materialsamp Continua vol 52 no 3 pp 187ndash196 2016
[38] Y Cao Z Zhou X Sun and C Gao ldquoCoverless informationhiding based on the molecular structure images of materialrdquoComputers Materials and Continua vol 54 no 2 pp 197ndash2072018
[39] Y LiuH Peng and JWang ldquoVerifiable diversity ranking searchover encrypted outsourced datardquo Cmc-Computers Materials ampContinua vol 55 no 1 pp 037ndash057 2018
[40] K Lin J Lu C-S Chen and J Zhou ldquoLearning compactbinary descriptors with unsupervised deep neural networksrdquo inProceedings of the 2016 IEEEConference onComputer Vision andPattern Recognition CVPR 2016 pp 1183ndash1192 USA July 2016
[41] T Do A Doan and N Cheung ldquoLearning to Hash with BinaryDeep Neural Networkrdquo in Computer Vision ndash ECCV 2016vol 9909 of Lecture Notes in Computer Science pp 219ndash234Springer International Publishing Cham 2016
[42] Rui Zhang Di Xiao and Yanting Chang ldquoA Novel ImageAuthentication with Tamper Localization and Self-Recovery inEncrypted Domain Based on Compressive Sensingrdquo Securityand Communication Networks vol 2018 Article ID 1591206 15pages 2018
[43] Xia ShuangKui and JianbinWu ldquoAModification-Free Steganog-raphy Method Based on Image Information Entropyrdquo Securityand Communication Networks vol 2018 Article ID 6256872 8pages 2018
[44] J Zhang B Qu and N Xiu ldquoSome projection-like methods forthe generalized Nash equilibriardquo Computational Optimizationand Applications vol 45 no 1 pp 89ndash109 2010
[45] Biao Qu and Jing Zhao ldquoMethods for Solving Generalized NashEquilibriumrdquo Journal of Applied Mathematics vol 2013 ArticleID 762165 6 pages 2013
[46] CWang CMa and J Zhou ldquoA new class of exact penalty func-tions and penalty algorithmsrdquo Journal of Global Optimizationvol 58 no 1 pp 51ndash73 2014
[47] Y Wang X Sun and F Meng ldquoOn the conditional andpartial trade credit policywith capital constraints A StackelbergModelrdquo Applied Mathematical Modelling vol 40 no 1 pp 1ndash182016
[48] S Lian and Y Duan ldquoSmoothing of the lower-order exactpenalty function for inequality constrained optimizationrdquo Jour-nal of Inequalities and Applications Paper No 185 12 pages2016
Chemical EngineeringInternational Journal of Antennas and
Propagation
International Journal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Navigation and Observation
International Journal of
Hindawi
wwwhindawicom Volume 2018
Advances in
Multimedia
Submit your manuscripts atwwwhindawicom
Security and Communication Networks 3
the visible units in the first layer are successfully applied tothe processing of massive image data
In [27] a deep neural network was developed to learnmultilevel nonlinear transformations mapping the originalimage to a compact binary hash code to support large-scaleimage retrieval for learning binary image representationsA deep hashing model is established under the three-layerconstraint on the deep network (1) the reconstruction errorbetween the original real-valued image feature vector andthe resulting binary code is minimized (2) each bit of thebinary code has a balance (3) all bits are independent of eachother Similar constraints were used in previous unsupervisedhash or binary codingmethods such as iterative quantization(ITQ) [28] In [27] a supervised version is called superviseddepth hash 3 in which a discriminative item containing twopairs of supervised information is added to the objectivefunction of the deep hash model
It is worth noting that in the training of sparse neuralnetworks in addition to the sparse similarity maintaininghash depth hash and supervised depth hash the pretrainingphase is not included Instead the hash codes are learnedfrom scratch using a set of training data However nopretraining can make the general hash code less efficient Inparticular the sparse similarity keep hash method is foundto be inferior to the existing supervised hashing methodie kernel-based supervised hashing (KSH) [29] in terms ofsearch accuracy on some image datasets [30] the deep hashmethod and its supervised version are slightly better than ITQand its supervised version CCA+ ITQ respectively [31] Notethat KSH ITQ and CCA+ITQ develop a shallow learningframework
One of the main purposes of deep learning is to learnthe robust and powerful representation of complex data Itis very natural to use deep learning to explore compact hashcodes which can be thought of as binary representation ofthe data The deployed CNN consists of three convolutionpools including rectified linear activation maximum poolmerging and local contrast normalization a standard fullyconnected layer and an output layer with softmax activationfunctions In [32] a newmethod called deep semantic sortinghashing is proposed to learn hash values therebymaintainingmultilevel semantic similarity between multilabel imagesThis method is combined with convolutional neural networkhashing method taking image pixels as input training depthCNN and jointly learning image feature representation andhash value by thismethodThe deployed CNN consists of fiveconvolution-pooling layers two fully connected layers and ahash layer (ie output layer)
Indexing massive amounts of multimedia data such asimages and videos is a natural application based on learninghashing In particular due to the well-known semanticdivide hashing methods have been extensively studied forimage search and retrieval as well as mobile product search[33 34] Although hashing techniques have been appliedto the active learning framework to cope with big dataapplications these hash algorithms are rarely applied toimage classification For the image recognition problem withmany categories the computational and memory overheadprimarily stems from the large number of classifiers to be
Table 1 A huge number of parameters need to be learned andstored with different datasets Considering a classification task withC different categories and D-dimensional feature representationeven the simplest linear models are comprised of DtimesC parameters
learned Table 1 show that the complexities can be high atthe stages of both training and deploying these classifiersInspired by [35 36] we proposed a combination classifierswith DNN and binarizing classifiers to solve some classicsecurity calculation problems [37ndash39] Different from otherdeep learning-hash methods that often entails heavy compu-tations by using a conventional classifier exemplified by K-NNand SVM ourmethod learns classifiers using binary hashcodes which are simultaneously learned from the trainingdata The combination classifiers can provide both imagefeatures and accelerate image classification and thus it makethe large-scale image recognition faster The advantages ofthe extending hashing techniques from fast image search toimage classification inspire us to apply them to deep learninghash framework
With the development of Internet technology the spread-ing form of illegal information on the Internet is changinggradually Traditionally the form of communication basedon word description has been transformed into a diversifiedformof communication based on video and imageThereforethe original keyword blocking web content grading blacklistrestriction and other filtering methods have not been able toblock illegal information dissemination At present networkillegal image recognition is mainly divided into the followingcategories
(1) Erotic Image Recognition The identification of onlinepornographic imagesmainly includesmethods based on skincolor feature extraction detection and judgment andmethodsbased on limb judgment The limb state is determinedby extracting the divided skin color information and theconnection feature of the human body posture and furtherdetermining whether the image transmitted on the networkcontains pornographic content
(2) Images Involving State Secrets or Military Secrets Therecognition of secret-related images mainly involvessteganography which hides secret information in imagevideo and other carriers for secret transmission At presentthe high-order statistical features based on modeling thecomplex correlation of image neighborhood become themainstream features in the field of steganalysis SRM(Spatial Rich Model) PSRM (Projection Speciation RichModel) and other models are based on such high-orderhigh-dimensional features and have achieved good detectionresults Steganalysis based on depth learning is a hotspot in
4 Security and Communication Networks
the field of information hiding in order to identify the illegalsecret-related images accurately and quickly
(3) Images Containing Antihuman Content such as TerroristViolenceThe identification of such images is mainly based onimage contrast techniques Image comparison includes tech-niques such as image feature extraction high-dimensionalspatial feature index establishment and similarity measureIt is a very worthwhile to study how to quickly comparethe massive network images to the illegal target images sothat the recall rate and the precision rate can be taken intoaccount
3 The Proposed Method
In this section we will present the notations as summarizedin Table 2 firstly The concept of deep learning stems fromthe field of artificial neural networks Deep learning is deepneural network learning and is a learning structure withmultiple hidden layers In the process of deep learningthe network is trained layer by layer Each layer of thelearning network extracts certain features and informationand takes the training result as a deeper input Finallythe entire network is fine-tuned with a top-down algo-rithm
Through deep learning complex function expressionscan be learned thereby completing the concept of high-levelabstraction from the underlying information It has beenwidely used in language understanding target recognitionand speech perception
Figure 1 shows that the proposed framework IDLHincludes three components The first component is prepro-cessing layer on the image dataset The second component isthe training self-coding network with a layer-by-layer greedylearning algorithm to obtain the feature expression functionof the image The third is hash layer which retrieves imagessimilar to the query image with compact binary codes andcategorizes the query one by the majority semantic labelwithin the hashing bucket
31 Preprocessing Since the depth learning algorithm usedin this paper is an unsupervised learning algorithm it canautomatically learn the deep features of the image fromthe original pixel information of the image Therefore theoriginal pixel value of the whole image can be directly used asinput data for the deep learning model In order to facilitatethe training of the network it is necessary to preprocess theimage Through preprocessing the image is simply scaledsample-by-sample mean-value reduction and whitening isprocessed to reduce the redundant information in the imageand facilitate the deep learning network for training andcalculation Preprocessing can be further grouped into threesuboperations
(1) Normalization Normalization can prevent neuron outputsaturation caused by excessive net input absolute value Weuse the sigmoid function to do normalization as shown asfollows
Table 2 Summary of notations
Symbol Definition119909119894 the gray value of the image pixels120583(119894) mean pixels
U an arbitrary orthogonal matrix and defines in theZCA whitening
J(Wb) the quantization loss between the learned binaryvalues and the real values
120582 a weight attenuation parameter119878119894 119894119905ℎ seta b two thresholdsD dimensionality of data pointsM K number of candidate imagesE the objective function
119909119894 = 1(1 + 119890minus119909119894) (1)
119909119894 is the gray value of the image pixels
(2) Pointwise Mean Reduction This process mainly is to getrid of the redundant information of the image the meanvalue is eliminated for each point of the image the averagebrightness of the image is removed and the DC componentof the data is eliminated Assuming that 119909(119894) isin 119877119899 is the grayvalue of each pixel of image I we use formulae (2) and (3) tozero-mean image
120583(119894) = 1119899119899sum119895=1
119909(119894)119895 (2)
119909(119894)119895 = 119909(119894)119895 minus 120583(119894) (3)
(3) Whitening Whitening is an important pretreatmentprocess its purpose is to reduce the redundancy of input dataso that the whitened input data has the following properties(i) low correlation between features (ii) all features having thesame variance then the formula is as formula (4)
In formula (4) the rotation matrix of 119909119903119900119905119894 is 119880119879119909119894Generally when 119909 is in interval [-11] 120576 asymp 10minus5
32 Training Stack Sparse Self-Encoding Network Stack self-coding neural network has strong expressive ability whichmainly benefits from its hierarchical feature representationThrough one level of feature learning we can learn thehierarchical structure between features Stack self-encodingneural network is a neural network model composed ofmultilayer sparse self-encoder that is the output of theformer self-encoder as the input of the latter self-encoder
In the training the original input 119909(119896) is used as inputto train the first self-encoded neural network At this pointfor each training sample 119909(119896) the output ℎ(119896)1 of the hidden
Security and Communication Networks 5
Hash Layer 2
Hamming Retrieval
Query Result
Weights
Codes
Training Dataset
Feature expression
function
Image Preprocess Layer Feature
expression function
Hash Layer 1
Binary Weights
Binary Codes
Feature migration
Query Image
Figure 1 Deep learning-hash retrieval framework IDLH consists of three main components (preprocessing layer deep neural networklayer and hash layer) The object of the first layer is simply scaled sample-by-sample mean-value reduction and whitening In the secondcomponent we develop a deep neural network to obtain the feature expression function of the image And the classifier weights and featurebinary codes are simultaneously learned in the last component-hash layer
layer can be obtained and the output of the hidden layer canbe used as the input of the second self-encoder to continuetraining the second self-encoder Then the output ℎ(119896)2 of thesecond hidden layer of the self-encoder can be obtained Theoutputℎ(119896)1 of the first hidden layer of the self-encoder is calleda first-order feature and the output ℎ(119896)2 of the second hiddenlayer of the self-encoder is called a second-order feature Inorder to classify the two-order feature ℎ(119896)2 can be used as theinput of Softmax regression
Figure 2 shows the flowchart of the proposed methodAnd there are mainly three processes (supervised pretrain-ing rough image retrieval and accurate image retrieval)The object of the first process is to transform the high-dimensional feature vector into a low-dimensional compacttwo value codes through hash function In the secondprocedure we pick out M candidate images by calculatingHamming distance In the third process we calculate theEuclidean distance between the candidate image and theimage to be retrieved and accurately extract K images fromthe M candidate images
Figure 3 shows a block diagram of a self-encoding neuralnetwork The stacking self-encoding network contains 3hidden layers (feature layers) The input layer inputs theoriginal data i into the first layer of the feature layer theoutput result of the former layer serves as the input of thenext layer and the output of the third layer serves as thefeature expression of the image In our method it is also usedas the input of the hash classifier and it is possible to usethe characteristics of the STD neural network to classify thefeatures
By using the matrix representation of the binary codesvectors and the output of the 3th layer of the network we usethe gradient descent method to solve the neural network
For a single sample (x y) the cost function is as shown in
(12 10038171003817100381710038171003817ℎ119882119887 (119909(119894)) minus 119910(119894)100381710038171003817100381710038172)]
+ 1205822119899119897minus1sum119897=1
119904119897sum119894=1
119904119897+1sum119895=1
(119882(119897)119895119894 )2
(6)
The first term 119869(119882 119887) represents the mean variance termThe second term aims to prevent the data from overfitting byreducing the magnitude of the weight 120582 is a weight attenua-tion parameter It is used to balance the relative importance ofmean square deviation terms and weight attenuation termsOur purpose is to minimize the quantization loss 119869(119882 119887)between the learned binary values and the real values of aninput image according to parameters119882 and 11988733 Hash Algorithm Retrieval The image retrieval methodbased on hash algorithmmaps the high-dimensional contentfeatures of images into Hamming space (binary space) andgenerates a low-dimensional hash sequence to represent apicture This method reduces the requirement of computermemory space for image retrieval system improves theretrieval speed and better adapts to the requirements of massimage retrieval
Inspired by [6 8] we use a set of hash functions to hashdata into different buckets After we do some hash mappingon the original image feature data we hope that the originaltwo adjacent feature data can be hash into the same bucketwith the same bucket number And then after hash mappingof all the data in the original feature set we can get a hashtable These original feature data sets are scattered into hashtable buckets and the data belonging to the same bucket isprobably adjacent to the original data However there is also
6 Security and Communication Networks
StartTwo valued by the
Sigmod function in Hash layer
Building hash pool and hash bucket
Calculation of Hamming distance
Select the top m candidate images
Calculation of Euclidean distance
Select the top kimages from m
candidate imagesas retrieval result
End
Supervised pre-training
Rough image retrieval
If itgt05Yes
1
No
0
If it less than a threshold
Yes
No
Accurate image retrieval
The featureof the image
to be retrieved
Self-learning network
The featureof the
training images
Binary processing
Binary codes for trainingimage
Binary codes for retrieving
image
Figure 2 Deep learning-hash retrieval flowchart IDLHmainly includes three processes (supervised pretraining rough image retrieval andaccurate image retrieval)
+1+1 +1
+1
Layer 0Input layer
Layer 1 Layer 2 Layer 3 Hash Layer
x1
x2
x3
x4
x5
ℎ(1)1
ℎ(1)2
ℎ(1)3
ℎ(1)4
ℎ(2)1
ℎ(2)2
ℎ(2)3
ℎ(2)4
ℎ(3)1
ℎ(3)2
ℎ(3)3
Output ℎ(3)k
from layer 3 is usedas binary code with bk = f(ℎ(3)
k)
Figure 3 Self-learning network based on stack self-encoding network The neurons labeled xi is the input of the neural network and ldquo+ 1rdquoare the offset nodes (intercept entries) of the neural network The layer 0 is the input layer of neural network and layer 3 is the output layerof neural network The middle layers of layer 0 layer to layer 3 are the hidden layers of neural network
a small probability in events that is the nonadjacent data ishash to the same barrel Set the hash function as the follows
Here119908119896 is the projection vector and 119887119896 is the correspond-ing intercept The code value generated by formula (7) isminus1 1 and we use the following formula to convert it intotwo value codes
119910119896 = 12 (1 + ℎ119896 (119909)) (8)
Given a sample point isin 119877119863 we can compute a K-bitbinary code 119910 for 119909 with formula (9) The hash functionperforms the mapping as ℎ119896 119877119863 997888rarr 119861
Here 119883 = 119909119899119873119899=1 isin 119877119863times119873 is the feature data matrix withpoints as columns Such a binary encoding process can also beviewed as mapping the original data point to a binary valuedspace namely Hamming space
34 SimilarityMeasure After obtaining the binary hash codeof the image it is necessary tomeasure similarity between theretrieved image and the library image in the Hamming spaceThe smaller the Hamming distance is the closer distancebetween the two data is and the degree of similarity is higherotherwise the two data similarity is lower
Here oplus is an XOR operation The two sets of 119910119894 and 119910119895represent the hash code of the search image feature and theimage library is mapped through the hash function The newimage features learned by the stack self-encoding network aregenerated by the hash function The storage structure of theimage feature vectors is shown in Table 3
As can be seen from Table 3 the hash code of the imageis related to the image ID and the image name one by one Inthe process of searching the image feature vector is obtainedthrough deep learning by a hash function the original data ismapped into a newdata space and a corresponding hash codeis obtained The hash code is used to calculate the Hammingdistance in the Hamming space as a measure of similaritybetween images Finally the storage structure of the imagefeature vector is used to find the corresponding image ID ofthe hash code and the output retrieval result is output to theuser
35 Image Secondary Search In the first-level search phasethe features learned from the deep learning network aremapped into the Hamming space using the hash function Inthe similarity measurement phase the traditional Euclideandistance is abandoned Measure the similarity betweenimages by comparing the Hamming distance between theimage features of the query image and the image of thelibrary image In order to further improve the accuracy ofretrieval without affecting the real-time performance we canretrieve the image by the second level retrieval These stepsare described in detail as follows After one level retrievalwe choose the K images with the most similarity in thefirst-level retrieval result and then calculate the Euclideandistance between the original feature vector of the K imagesand the original feature vector of the query imageThe resultsobtained as the similarity measure of the images and outputthe retrieval result that has been ranked from the high andlow with the similarity distance
Although theHash algorithmmaps the high-dimensionalfeature vectors of the image into a hash-coded form theproblem of ldquodimensional disastersrdquo is solved and the retrievalefficiency is greatly accelerated However when the similaritycomparison is performed the Hamming distances of theimage features are simply compared using the results of theprimary search and occasionally undesirable results maystill appear on the search results If we want to increasethe accuracy of the search we must increase the hash codelength However excessively long codes will increase theamount of calculations increase the memory burden andreduce the real-time nature of retrieval failing to achieve
the goal of reducing the size of data In order to solve thisproblem keep the retrieval efficiency and further improvethe retrieval accuracy we propose a search strategy forsecondary retrieval the specific steps of which are as follows
Step 1 Through the first-level search in the Hamming spacethe similarity degree of the images is sorted and the top Ksorting images are selected
Step 2 For the 119896 images in Step 1 calculate the Euclideandistance one by one from its original image feature vector tothe image feature vector of the query image
Step 3 The Euclidean distance calculated in Step 2 is sortedThe smaller the calculated value is the higher similaritybetween images is and the similarity is sorted from high tolow and output as the final search result
In the second search it is necessary to pay attention tothe selection of the 119870 value although the larger the 119870 valueis the better the search effect is but accordingly the longerthe time is consumed Therefore it is necessary to combinevarious factors to select the appropriate119870 value
4 Experimental and Performance Analysis
In this section we thoroughly compare the proposedapproach with the improved deep learning hash retrievalmethods on several benchmark datasets Through a series ofexperiments the effectiveness and feasibility of the proposedalgorithm are verified
41 Database Two mostly used databases in the recent deeplearning hash works are taken into our evaluation Thetwo image libraries are derived from the CIFAR-10[11] coreexperimental image library dataset and theCaltech 256 imagelibrary dataset
CIFAR-10 dataset contains 10 object categories and eachclass consists of 6000 images resulting in a total of 60000images The dataset is split into training and test sets whichare averagely divided into 10 object classes The Caltech 256image library dataset contains 29780 color images which aregrouped intro 256 classes
First test the CIFAR-10 image dataset There are a totalof 50000 training samples which are used for training on thedeep learning networkThe remaining 10000 images are usedas test samples And then we randomly select 50 images fromdatabase as the query images For theHidden Image Retrievalalgorithm based on deep learning mentioned in this paperthe image pixel data is directly used as input while for otheralgorithms the 512-dimensional GIST feature is used as thefeature expression of the image Note quantization all imagesinto 32lowast32 sizes before experiment
For the Caltech 256 image a total of 256 classes areincluded and each class contains at least 70 images There-fore 70 images of each class a total of 17920 images arerandomly selected and are used as training images Theremaining images are used as test samples In addition all ofthe imagesrsquo size is set to 64lowast64 again when training
8 Security and Communication Networks
42 Evaluation Metrics We measure the performance ofcompared methods using Precision-recall and Average-Retrieval Precision (ARP) curves Precision is the ratio ofthe correct number of images m in the search result to thenumber k of all returned images The formula is as follows
precision = 119898119896 times 100 (12)
Recall is the ratio of the correct number of imagesm in thesearch results to the number g of images in the image libraryThe formula is as follows
recall = 119898119902 times 100 (13)
Assume that the search result of the query image 119894 is119861119894 and 119860 119894 means that the category is the same between thequery image and the return image then the accuracy rate forthe image query result 119875(119894) can be defined by the followingformula
119875 (119894) = |119860 (119894) cap 119861 (119894)||119861 (119894)| (14)
Average-Retrieval Precision (ARP) the average value ofall the images in the same class as the retrieval rate obtainedfrom the retrieval image is defined as follows
Here 119868119863119898 is the category index number of the image119898 isthe category index119873 is the number of imageswhose categoryis 119868119863119898 and 119894119889(119894) is the category index number of the queryimage
43 Performance Analysis In the proposed algorithm IDLHthe length of the hash sequence and the depth of the hiddenlayer in the deep learning network are two key parametersWhen the hash sequence length is small different featurevectors can easily be mapped into the same hash sequence sothe retrieval accuracy is low However if the hash sequence istoo long a large storage space is required and a long time isconsumed which reduces the real-time performance For thenumber of hidden layers the number of layers in the hiddenlayer is too small which is not conducive to learning strongimage features However if the depth of the hidden layer istoo large the difficulty of training is increased In order toverify the effectiveness and feasibility of our algorithm weconducted the following experiments
(1) Results on CIFAR-10 Dataset Figure 4(a) shows the searchresults of the Average-Retrieval Precision using our proposedalgorithm IDLH compared with the LSH algorithm [3] andother three deep learning algorithms the DH algorithm [21]the DeepBit algorithm [40] and the UH-BDNN algorithm[41] on CIFAR-10 dataset with 8 16 32 48 64 96 128and 256 bits Figure 4(b) shows the Precision-recall curveunder 48-bit encoding It can be seen that the algorithm hasa higher precision than the other hashing algorithms with
the same recall rate However the advantage is not obviousand the average accuracy is slightly higher than other hashalgorithms
In order to overcome the above defects we use deeplearning to perform hash mapping on image features per-form hash encoding of different bits on the same featurecalculate the Precision-recall of the search results under thecondition of different coded bits and determine the impactof the encoding length on the retrieval results
As Figure 5 shows with the increase in the number ofcoded bits the Precision-recall is continuously increasingWith the increase in the number of coded bits the imageis better expressed However after the number of coded bitsreaches 64 even if the number of coded bits increases theaverage accuracy rate increases relatively slowly Because theinformation of the tiny image is relatively simple when thenumber of encoded bits reaches 64 bits a relatively goodimage expression has been obtained and the performanceof the algorithm has basically stabilized At this time despiteincrease in the number of encoding bits it is not very helpfulto improve the accuracy rate
In addition we want to test the influence of the numberof hidden layers in the deep learning network on the retrievalresult by changing the number of hidden layers
Figure 6 shows the effect of deep learning networks onexperimental results in the case of different hidden layernumbers
It can be seen that deeper networks do not have muchimprovement in performance which is different from theexpectation that more hidden layers will help learn strongerimage features Since the image library data used in theexperiment is a tiny image library relatively good image char-acteristics can be learned using a deep learning network withfewer layers However if the image library is replaced witha more colorful image the deep neural network can acquiremore detailed image features and the deepening of thelearning network will greatly help the study of image features
(2) Results on Caltech 256 Image Data Set Figure 7(a)shows the results of the Average-Retrieval Precision resultswhen the number of coded bits is different Compared withthe black and white image library the proposed algorithmembodies the advantage of image feature learning and leadsthe Average-Retrieval Precision to other hash retrieval algo-rithms In Figure 7(b) we can also see that the algorithmproposed in this paper has a higher Precision-recall thanother algorithms under the same recall rate and it has bettersearch performance
As shown in Figure 8 as the number of coded bitsincreases the precision rate increases with the same recallrate This feature of the color image library is more pro-nounced than the black and white image library Becausethe color image contains more information more codingis needed to express it and the increase of encoding helpsto learn the features of the image The experimental resultsalso show that the deep learning network has learned moreexcellent image features
In Figure 9 the precision rate is significantly improvedby the increase in the number of hidden layers in the
Security and Communication Networks 9
IDLHLSHDHDeepBitUH-BDNN
16 32 48 64 96 128 2568Bits
0
01
02
03
04
05
06
07
08Av
erag
e Ret
rieva
l Pre
cisio
n
(a) Average-Retrieval Precision
IDLHLSHDHDeepBitUH-BDNN
01 02 03 04 05 06 07 08 09 100Recall
0
01
02
03
04
05
06
07
08
09
Prec
ision
(b) Precision-recall at 48 bits
Figure 4 Five kinds of algorithm retrieval performance comparison on CIFAR-10 dataset
8 bits16 bits32 bits48 bits64 bits128 bits
01 02 03 04 05 06 07 08 09 10Recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
Figure 5 Precision-recall curves with lengths
color image library Caltech 256 This is because theinformation contained in a more colorful image is morecomplex Adding a hidden layer can learn more detailsof the image and help improve the accuracy of thesearch
Next we tested the performance of secondary imageretrieval The value of k in the secondary search is 20 andthe experimental results are shown in Figure 10 As can
Level 1Level 2Level 3Level 4
01 02 03 04 05 06 07 08 09 1000Recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
Figure 6 Precision-recall curves with different code differenthidden layers
be seen from the results secondary retrieval can effectivelyimprove the retrieval accuracy when the number of codedbits is small However with the increase in the number ofencoding bits the results of the secondary search and theaccuracy of the primary search are not much different Thisis because the shorter the hash sequence is the easier thefeature vectors with different original features are mappedto the same hash code In order to make up for the errors
10 Security and Communication Networks
IDLHLSHDHDeepBitUH-BDNN
16 32 48 64 96 128 2568Bits
0
01
02
03
04
05
06
07
08
09
1Av
erag
e-Re
trie
val P
reci
sion
(a) Average-Retrieval Precision
IDLHLSHDHDeepBitUH-BDNN
01 02 03 04 05 06 07 08 09 1000Precision-recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
(b) Precision-recall at 48 bits
Figure 7 Five kinds of algorithm retrieval performance comparison on Caltech 256 set
8 bits16 bits32 bits48 bits64 bits128 bits
01 02 03 04 05 06 07 08 09 10Recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
Figure 8 Precision-recall curves with different code lengths
caused by the short hash code it is necessary to performsecondary searchWhen the number of encoding bits is smalla secondary retrieval method is used in IDLH and the searchaccuracy rate can be improved at the expense of a small searchspeed
Level 1Level 2Level 3Level 4
01 02 03 04 05 06 07 08 09 10Recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
Figure 9 Precision-recall curves with different hidden layers
5 Conclusion
With the rapid development of data storage and digital pro-cess more and more digital information is transformed andtransmitted over the Internet day by day which brings peoplea series of security problems as well as convenience [42] Theresearches on digital image security that is image encryptionimage data hiding and image authentication become moreimportant than ever [38 39] The most essential problem of
Security and Communication Networks 11
First retrievalSecond retrieval
16 32 48 64 96 128 2568Bits
07
08
09
1
Aver
age-
Retr
ieva
l Pre
cisio
n
Figure 10 Average-Retrieval Precision with first and secondretrieval
image recognition is to extract robust features The qualityof feature extraction is directly related to the effect of recog-nition so most of the previous work on image recognitionis spent on artificial design features [43] In recent yearsthe emergence of deep learning technology has changed thestatus of artificial design classification characteristics Deeplearning technology simulates the mechanism of humanvisual system information classification processing from themost primitive image pixels to lower edge features then tothe target components that are combined on the edge andfinally to the whole target depth learning can be combinedby layer by layer The high-level feature is the combination oflow level features From low level to high level features aremore and more abstract and show semantics more and moreFrom the underlying features to the combination of high-levelfeatures it is the depth of learning that is done by itself Itdoes not require manual intervention Compared with thecharacteristics of the artificial design this combination offeatures can be closer to the semantic expression
In terms of illegal image retrieval the traditional recog-nition method should establish a recognition model foreach type of recognition task In the actual application arecognition model needs a recognition server If there aremany identification tasks the cost is too high We used thedeep neural network to recognize the illegal image it onlyneeds to collect the samples of every kind of illegal imageand participate in the training of the deep neural networkFinally a multiclassification recognition model is trainedWhen classifying unknown samples deep neural networkaccounting calculates the probability that the image belongsto each class
We all know that in the image detection process theaccuracy and recall rate are mutually influential Ideally bothmust be high but in general the accuracy is high and the
recall rate is low the recall rate is high and the accuracy islow For image retrieval we need to improve the accuracyunder the condition of guaranteeing the recall rate Forimage disease surveillance and anti-illegal images we need toenhance the recall under the condition of ensuring accuracyTherefore in different application scenarios in order toachieve a balance between accuracy and recall perhaps somegame theory (such as Nash Equilibrium [44 45]) and penaltyfunction [46ndash48] can provide related optimization solutions
In this paper we proposed an improved deep-learning-hashing approach IDLH which optimized over two majorimage retrieval process
(a) In the feature extraction process the self-encodednetwork of the look-ahead type is trained by using unlabeledimage data and the expression of robust image features islearned This unlabeled learning method does not requireimage library labeling and reduces the requirements for theimage library At the same time it also takes advantage of thedeep learning networks strong learning ability and obtainsbetter image feature expression than ordinary algorithms
(b) On the index structure a secondary search is pro-posed which further increases the accuracy of the search atthe expense of very little retrieval time
Through experiments the algorithm proposed in thispaper is compared with other classic hashing algorithms onmultiple evaluation indicators Firstly we tested the learningnetworks of different code lengths and depths in order totest their effect on the retrieval system and then tested theperformance of the secondary search Through the above-mentioned series of experiments for different parameters theeffectiveness of the improved deep learning hash retrievalalgorithm proposed in this paper is verified and throughthe experimental data the good retrieval results are provedIn addition the proposed deep hashing training strategycan also be potentially applied to other hashing problemsinvolving data similarity computation
Data Availability
The data used to support the findings of this study areincluded within the article
Conflicts of Interest
The authors declare that they have no conflicts of interest
Acknowledgments
The work was funded by the National Natural ScienceFoundation of China (Grants nos 61206138 and 61373016)
References
[1] R Datta D Joshi J Li and J Z Wang ldquoImage retrieval ideasinfluences and trends of the new agerdquoACMComputing Surveysvol 40 no 2 article 5 2008
[2] G Shakhnarovich T Darrell and P Indyk Nearest-NeighborMethods in Learning andVisionTheory and PracticeMITPressCambridge MA USA 2006
12 Security and Communication Networks
[3] A Gionis P Indyk and R Motwani ldquoSimilarity search in highdimensions via hashingrdquo in 25th Int Conf pp 518ndash529 1999
[4] Z Pan J Lei Y Zhang and F L Wang ldquoAdaptive fractional-Pixel motion estimation skipped algorithm for efficient HEVCmotion estimationrdquoACMTransactions onMultimedia Comput-ing Communications and Applications (TOMM) vol 14 no 1pp 1ndash19 2018
[5] G-L Tian M Wang and L Song ldquoVariable selection in thehigh-dimensional continuous generalized linear model withcurrent status datardquo Journal of Applied Statistics vol 41 no 3pp 467ndash483 2014
[6] M Datar N Immorlica P Indyk and V S Mirrokni ldquoLocality-sensitive hashing scheme based on p-stable distributionsrdquo inProceedings of the 20th Annual Symposium on ComputationalGeometry (SCG rsquo04) pp 253ndash262 ACM June 2004
[7] B Kulis P Jain and K Grauman ldquoFast similarity search forlearned metricsrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 31 no 12 pp 2143ndash2157 2009
[8] M Raginsky and S Lazebnik ldquoLocality-sensitive binary codesfrom shift-invariant kernelsrdquo in Proceedings of the 23rd AnnualConference on Neural Information Processing Systems NIPS2009 pp 1509ndash1517 Canada December 2009
[9] L Qi X Zhang W Dou and Q Ni ldquoA distributed locality-sensitive hashing-based approach for cloud service recommen-dation from multi-source datardquo IEEE Journal on Selected Areasin Communications vol 35 no 11 pp 2616ndash2624 2017
[10] M A Carreira-Perpinan and R Raziperchikolaei ldquoHashingwith binary autoencodersrdquo in Proceedings of the IEEE Confer-ence on Computer Vision and Pattern Recognition CVPR 2015pp 557ndash566 USA June 2015
[11] J Wang S Kumar and S-F Chang ldquoSemi-supervised hashingfor large-scale searchrdquo IEEE Transactions on Pattern Analysisand Machine Intelligence vol 34 no 12 pp 2393ndash2406 2012
[12] Y Gong S Lazebnik and A Gordo ldquoIterative quantizationa Procrustean approach to learning binary codes for large-scale image retrievalrdquo in Proceedings of the IEEE Conference onComputer Vision and Pattern Recognition (CVPR rsquo11) pp 2916ndash2929 June 2011
[13] W Kong and W J Li ldquoIsotropic hashingrdquo NIPS vol 25 2012[14] M Norouzi and D J Fleet ldquoMinimal loss hashing for compact
binary codesrdquo in Proceedings of the 28th International Confer-ence on Machine Learning ICML 2011 pp 353ndash360 USA July2011
[15] J Wang W Liu A X Sun and Y-G Jiang ldquoLearning hashcodes with listwise supervisionrdquo in Proceedings of the 2013 14thIEEE International Conference on Computer Vision ICCV 2013pp 3032ndash3039 Australia December 2013
[16] G Lin C Shen Q Shi A Van Den Hengel and D Suter ldquoFastsupervised hashing with decision trees for high-dimensionaldatardquo in Proceedings of 27th IEEE Conference on ComputerVision and Pattern Recognition CVPRrsquo pp 1971ndash1978 USA2014
[17] Y Gong S Kumar H A Rowley and S Lazebnik ldquoLearningbinary codes for high-dimensional data using bilinear projec-tionsrdquo in Proceedings of the 26th IEEE Conference on ComputerVision and Pattern Recognition CVPR 2013 pp 484ndash491 USAJune 2013
[18] W Liu J Wang Y Mu and S Kumar ldquoCompact hyperplanehashing with bilinear functionsrdquo in The 29th InternationalConference on Machine Learning (ICML12) pp 467ndash474 2012
[19] Y Weiss A Torralba and R Fergus ldquoSpectral hashingrdquo inProceedings of the 22nd Annual Conference on Neural Informa-tion Processing Systems (NIPS rsquo08) pp 1753ndash1760 VancouverCanada December 2008
[20] W Liu J Wang S Kumar and S F Chang ldquoHashing withgraphsrdquo inThe 28th international conference on machine learn-ing (ICML11) 2011
[21] F Shen X Zhou Y Yang J Song H T Shen and D Tao ldquoA fastoptimization method for general binary code learningrdquo IEEETransactions on Image Processing vol 25 no 12 pp 5610ndash56212016
[22] F Shen W Liu S Zhang Y Yang and H T Shen ldquoLearningbinary codes for maximum inner product searchrdquo in Proceed-ings of the 15th IEEE International Conference on ComputerVision ICCV 2015 pp 4148ndash4156 Chile December 2015
[23] A Krizhevsky I Sutskever andG EHinton ldquoImagenet classifi-cation with deep convolutional neural networksrdquo in Proceedingsof the 26th Annual Conference on Neural Information ProcessingSystems (NIPS rsquo12) pp 1097ndash1105 Lake Tahoe Nev USADecember 2012
[24] G E Hinton and R R Salakhutdinov ldquoReducing the dimen-sionality of data with neural networksrdquo The American Associa-tion for the Advancement of Science Science vol 313 no 5786pp 504ndash507 2006
[25] A Torralba R Fergus and Y Weiss ldquoSmall codes and largeimage databases for recognitionrdquo in Proceedings of the IEEEComputer Society Conference on Computer Vision and PatternRecognition (CVPR rsquo08) pp 1ndash8 2008
[26] R Salakhutdinov andG Hinton ldquoLearning a nonlinear embed-ding by preserving class neighbourhood structurerdquo Journal ofMachine Learning Research vol 2 pp 412ndash419 2007
[27] V E Liong J Lu GWang P Moulin and J Zhou ldquoDeep hash-ing for compact binary codes learningrdquo in Proceedings of theIEEE Conference on Computer Vision and Pattern RecognitionCVPR 2015 pp 2475ndash2483 USA June 2015
[28] Y Gong S Lazebnik A Gordo and F Perronnin ldquoIterativequantization A procrustean approach to learning binary codesfor large-scale image retrievalrdquo IEEE Transactions on PatternAnalysis andMachine Intelligence vol 35 no 12 pp 2916ndash29292013
[29] W Liu J Wang R Ji Y-G Jiang and S-F Chang ldquoSupervisedhashing with kernelsrdquo in Proceedings of the IEEE Conference onComputer Vision and Pattern Recognition (CVPR rsquo12) pp 2074ndash2081 Providence RI USA June 2012
[30] J Masci A Bronstein M Bronstein and P SprechmannldquoSparse similarity-preserving hashingrdquo in Int Conf LearnRepresent pp 1ndash13 2014
[31] B Kulis and K Grauman ldquoKernelized locality-sensitive hash-ingrdquo IEEE Transactions on Pattern Analysis and Machine Intel-ligence vol 34 no 6 pp 1092ndash1104 2012
[32] F Zhao YHuang LWang and T Tan ldquoDeep semantic rankingbased hashing for multi-label image retrievalrdquo in Proceedings ofIEEE Conference on Computer Vision and Pattern RecognitionCVPR 2015 pp 1556ndash1564 June 2015
[33] G Cheng C Yang X Yao L Guo and J Han ldquoWhenDeep Learning Meets Metric Learning Remote Sensing ImageScene Classification via Learning Discriminative CNNsrdquo IEEETransactions on Geoscience and Remote Sensing pp 1ndash11
[34] J He J Feng X Liu et al ldquoMobile product search with Bag ofHash Bits and boundary rerankingrdquo in Proceedings of the 2012IEEE Conference on Computer Vision and Pattern RecognitionCVPR 2012 pp 3005ndash3012 USA June 2012
Security and Communication Networks 13
[35] F Shen Y Mu Y Yang et al ldquoClassification by retrievalBinarizing data and classifiersrdquo in Proceedings of the 40thInternational ACM SIGIR Conference on Research and Develop-ment in Information Retrieval SIGIR 2017 pp 595ndash604 JapanAugust 2017
[36] P Li S Zhao andR Zhang ldquoA cluster analysis selection strategyfor supersaturated designsrdquo Computational Statistics amp DataAnalysis vol 54 no 6 pp 1605ndash1612 2010
[37] A Pradeep S Mridula and P Mohanan ldquoHigh securityidentity tags using spiral resonatorsrdquo Cmc-Computers Materialsamp Continua vol 52 no 3 pp 187ndash196 2016
[38] Y Cao Z Zhou X Sun and C Gao ldquoCoverless informationhiding based on the molecular structure images of materialrdquoComputers Materials and Continua vol 54 no 2 pp 197ndash2072018
[39] Y LiuH Peng and JWang ldquoVerifiable diversity ranking searchover encrypted outsourced datardquo Cmc-Computers Materials ampContinua vol 55 no 1 pp 037ndash057 2018
[40] K Lin J Lu C-S Chen and J Zhou ldquoLearning compactbinary descriptors with unsupervised deep neural networksrdquo inProceedings of the 2016 IEEEConference onComputer Vision andPattern Recognition CVPR 2016 pp 1183ndash1192 USA July 2016
[41] T Do A Doan and N Cheung ldquoLearning to Hash with BinaryDeep Neural Networkrdquo in Computer Vision ndash ECCV 2016vol 9909 of Lecture Notes in Computer Science pp 219ndash234Springer International Publishing Cham 2016
[42] Rui Zhang Di Xiao and Yanting Chang ldquoA Novel ImageAuthentication with Tamper Localization and Self-Recovery inEncrypted Domain Based on Compressive Sensingrdquo Securityand Communication Networks vol 2018 Article ID 1591206 15pages 2018
[43] Xia ShuangKui and JianbinWu ldquoAModification-Free Steganog-raphy Method Based on Image Information Entropyrdquo Securityand Communication Networks vol 2018 Article ID 6256872 8pages 2018
[44] J Zhang B Qu and N Xiu ldquoSome projection-like methods forthe generalized Nash equilibriardquo Computational Optimizationand Applications vol 45 no 1 pp 89ndash109 2010
[45] Biao Qu and Jing Zhao ldquoMethods for Solving Generalized NashEquilibriumrdquo Journal of Applied Mathematics vol 2013 ArticleID 762165 6 pages 2013
[46] CWang CMa and J Zhou ldquoA new class of exact penalty func-tions and penalty algorithmsrdquo Journal of Global Optimizationvol 58 no 1 pp 51ndash73 2014
[47] Y Wang X Sun and F Meng ldquoOn the conditional andpartial trade credit policywith capital constraints A StackelbergModelrdquo Applied Mathematical Modelling vol 40 no 1 pp 1ndash182016
[48] S Lian and Y Duan ldquoSmoothing of the lower-order exactpenalty function for inequality constrained optimizationrdquo Jour-nal of Inequalities and Applications Paper No 185 12 pages2016
Chemical EngineeringInternational Journal of Antennas and
Propagation
International Journal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Navigation and Observation
International Journal of
Hindawi
wwwhindawicom Volume 2018
Advances in
Multimedia
Submit your manuscripts atwwwhindawicom
4 Security and Communication Networks
the field of information hiding in order to identify the illegalsecret-related images accurately and quickly
(3) Images Containing Antihuman Content such as TerroristViolenceThe identification of such images is mainly based onimage contrast techniques Image comparison includes tech-niques such as image feature extraction high-dimensionalspatial feature index establishment and similarity measureIt is a very worthwhile to study how to quickly comparethe massive network images to the illegal target images sothat the recall rate and the precision rate can be taken intoaccount
3 The Proposed Method
In this section we will present the notations as summarizedin Table 2 firstly The concept of deep learning stems fromthe field of artificial neural networks Deep learning is deepneural network learning and is a learning structure withmultiple hidden layers In the process of deep learningthe network is trained layer by layer Each layer of thelearning network extracts certain features and informationand takes the training result as a deeper input Finallythe entire network is fine-tuned with a top-down algo-rithm
Through deep learning complex function expressionscan be learned thereby completing the concept of high-levelabstraction from the underlying information It has beenwidely used in language understanding target recognitionand speech perception
Figure 1 shows that the proposed framework IDLHincludes three components The first component is prepro-cessing layer on the image dataset The second component isthe training self-coding network with a layer-by-layer greedylearning algorithm to obtain the feature expression functionof the image The third is hash layer which retrieves imagessimilar to the query image with compact binary codes andcategorizes the query one by the majority semantic labelwithin the hashing bucket
31 Preprocessing Since the depth learning algorithm usedin this paper is an unsupervised learning algorithm it canautomatically learn the deep features of the image fromthe original pixel information of the image Therefore theoriginal pixel value of the whole image can be directly used asinput data for the deep learning model In order to facilitatethe training of the network it is necessary to preprocess theimage Through preprocessing the image is simply scaledsample-by-sample mean-value reduction and whitening isprocessed to reduce the redundant information in the imageand facilitate the deep learning network for training andcalculation Preprocessing can be further grouped into threesuboperations
(1) Normalization Normalization can prevent neuron outputsaturation caused by excessive net input absolute value Weuse the sigmoid function to do normalization as shown asfollows
Table 2 Summary of notations
Symbol Definition119909119894 the gray value of the image pixels120583(119894) mean pixels
U an arbitrary orthogonal matrix and defines in theZCA whitening
J(Wb) the quantization loss between the learned binaryvalues and the real values
120582 a weight attenuation parameter119878119894 119894119905ℎ seta b two thresholdsD dimensionality of data pointsM K number of candidate imagesE the objective function
119909119894 = 1(1 + 119890minus119909119894) (1)
119909119894 is the gray value of the image pixels
(2) Pointwise Mean Reduction This process mainly is to getrid of the redundant information of the image the meanvalue is eliminated for each point of the image the averagebrightness of the image is removed and the DC componentof the data is eliminated Assuming that 119909(119894) isin 119877119899 is the grayvalue of each pixel of image I we use formulae (2) and (3) tozero-mean image
120583(119894) = 1119899119899sum119895=1
119909(119894)119895 (2)
119909(119894)119895 = 119909(119894)119895 minus 120583(119894) (3)
(3) Whitening Whitening is an important pretreatmentprocess its purpose is to reduce the redundancy of input dataso that the whitened input data has the following properties(i) low correlation between features (ii) all features having thesame variance then the formula is as formula (4)
In formula (4) the rotation matrix of 119909119903119900119905119894 is 119880119879119909119894Generally when 119909 is in interval [-11] 120576 asymp 10minus5
32 Training Stack Sparse Self-Encoding Network Stack self-coding neural network has strong expressive ability whichmainly benefits from its hierarchical feature representationThrough one level of feature learning we can learn thehierarchical structure between features Stack self-encodingneural network is a neural network model composed ofmultilayer sparse self-encoder that is the output of theformer self-encoder as the input of the latter self-encoder
In the training the original input 119909(119896) is used as inputto train the first self-encoded neural network At this pointfor each training sample 119909(119896) the output ℎ(119896)1 of the hidden
Security and Communication Networks 5
Hash Layer 2
Hamming Retrieval
Query Result
Weights
Codes
Training Dataset
Feature expression
function
Image Preprocess Layer Feature
expression function
Hash Layer 1
Binary Weights
Binary Codes
Feature migration
Query Image
Figure 1 Deep learning-hash retrieval framework IDLH consists of three main components (preprocessing layer deep neural networklayer and hash layer) The object of the first layer is simply scaled sample-by-sample mean-value reduction and whitening In the secondcomponent we develop a deep neural network to obtain the feature expression function of the image And the classifier weights and featurebinary codes are simultaneously learned in the last component-hash layer
layer can be obtained and the output of the hidden layer canbe used as the input of the second self-encoder to continuetraining the second self-encoder Then the output ℎ(119896)2 of thesecond hidden layer of the self-encoder can be obtained Theoutputℎ(119896)1 of the first hidden layer of the self-encoder is calleda first-order feature and the output ℎ(119896)2 of the second hiddenlayer of the self-encoder is called a second-order feature Inorder to classify the two-order feature ℎ(119896)2 can be used as theinput of Softmax regression
Figure 2 shows the flowchart of the proposed methodAnd there are mainly three processes (supervised pretrain-ing rough image retrieval and accurate image retrieval)The object of the first process is to transform the high-dimensional feature vector into a low-dimensional compacttwo value codes through hash function In the secondprocedure we pick out M candidate images by calculatingHamming distance In the third process we calculate theEuclidean distance between the candidate image and theimage to be retrieved and accurately extract K images fromthe M candidate images
Figure 3 shows a block diagram of a self-encoding neuralnetwork The stacking self-encoding network contains 3hidden layers (feature layers) The input layer inputs theoriginal data i into the first layer of the feature layer theoutput result of the former layer serves as the input of thenext layer and the output of the third layer serves as thefeature expression of the image In our method it is also usedas the input of the hash classifier and it is possible to usethe characteristics of the STD neural network to classify thefeatures
By using the matrix representation of the binary codesvectors and the output of the 3th layer of the network we usethe gradient descent method to solve the neural network
For a single sample (x y) the cost function is as shown in
(12 10038171003817100381710038171003817ℎ119882119887 (119909(119894)) minus 119910(119894)100381710038171003817100381710038172)]
+ 1205822119899119897minus1sum119897=1
119904119897sum119894=1
119904119897+1sum119895=1
(119882(119897)119895119894 )2
(6)
The first term 119869(119882 119887) represents the mean variance termThe second term aims to prevent the data from overfitting byreducing the magnitude of the weight 120582 is a weight attenua-tion parameter It is used to balance the relative importance ofmean square deviation terms and weight attenuation termsOur purpose is to minimize the quantization loss 119869(119882 119887)between the learned binary values and the real values of aninput image according to parameters119882 and 11988733 Hash Algorithm Retrieval The image retrieval methodbased on hash algorithmmaps the high-dimensional contentfeatures of images into Hamming space (binary space) andgenerates a low-dimensional hash sequence to represent apicture This method reduces the requirement of computermemory space for image retrieval system improves theretrieval speed and better adapts to the requirements of massimage retrieval
Inspired by [6 8] we use a set of hash functions to hashdata into different buckets After we do some hash mappingon the original image feature data we hope that the originaltwo adjacent feature data can be hash into the same bucketwith the same bucket number And then after hash mappingof all the data in the original feature set we can get a hashtable These original feature data sets are scattered into hashtable buckets and the data belonging to the same bucket isprobably adjacent to the original data However there is also
6 Security and Communication Networks
StartTwo valued by the
Sigmod function in Hash layer
Building hash pool and hash bucket
Calculation of Hamming distance
Select the top m candidate images
Calculation of Euclidean distance
Select the top kimages from m
candidate imagesas retrieval result
End
Supervised pre-training
Rough image retrieval
If itgt05Yes
1
No
0
If it less than a threshold
Yes
No
Accurate image retrieval
The featureof the image
to be retrieved
Self-learning network
The featureof the
training images
Binary processing
Binary codes for trainingimage
Binary codes for retrieving
image
Figure 2 Deep learning-hash retrieval flowchart IDLHmainly includes three processes (supervised pretraining rough image retrieval andaccurate image retrieval)
+1+1 +1
+1
Layer 0Input layer
Layer 1 Layer 2 Layer 3 Hash Layer
x1
x2
x3
x4
x5
ℎ(1)1
ℎ(1)2
ℎ(1)3
ℎ(1)4
ℎ(2)1
ℎ(2)2
ℎ(2)3
ℎ(2)4
ℎ(3)1
ℎ(3)2
ℎ(3)3
Output ℎ(3)k
from layer 3 is usedas binary code with bk = f(ℎ(3)
k)
Figure 3 Self-learning network based on stack self-encoding network The neurons labeled xi is the input of the neural network and ldquo+ 1rdquoare the offset nodes (intercept entries) of the neural network The layer 0 is the input layer of neural network and layer 3 is the output layerof neural network The middle layers of layer 0 layer to layer 3 are the hidden layers of neural network
a small probability in events that is the nonadjacent data ishash to the same barrel Set the hash function as the follows
Here119908119896 is the projection vector and 119887119896 is the correspond-ing intercept The code value generated by formula (7) isminus1 1 and we use the following formula to convert it intotwo value codes
119910119896 = 12 (1 + ℎ119896 (119909)) (8)
Given a sample point isin 119877119863 we can compute a K-bitbinary code 119910 for 119909 with formula (9) The hash functionperforms the mapping as ℎ119896 119877119863 997888rarr 119861
Here 119883 = 119909119899119873119899=1 isin 119877119863times119873 is the feature data matrix withpoints as columns Such a binary encoding process can also beviewed as mapping the original data point to a binary valuedspace namely Hamming space
34 SimilarityMeasure After obtaining the binary hash codeof the image it is necessary tomeasure similarity between theretrieved image and the library image in the Hamming spaceThe smaller the Hamming distance is the closer distancebetween the two data is and the degree of similarity is higherotherwise the two data similarity is lower
Here oplus is an XOR operation The two sets of 119910119894 and 119910119895represent the hash code of the search image feature and theimage library is mapped through the hash function The newimage features learned by the stack self-encoding network aregenerated by the hash function The storage structure of theimage feature vectors is shown in Table 3
As can be seen from Table 3 the hash code of the imageis related to the image ID and the image name one by one Inthe process of searching the image feature vector is obtainedthrough deep learning by a hash function the original data ismapped into a newdata space and a corresponding hash codeis obtained The hash code is used to calculate the Hammingdistance in the Hamming space as a measure of similaritybetween images Finally the storage structure of the imagefeature vector is used to find the corresponding image ID ofthe hash code and the output retrieval result is output to theuser
35 Image Secondary Search In the first-level search phasethe features learned from the deep learning network aremapped into the Hamming space using the hash function Inthe similarity measurement phase the traditional Euclideandistance is abandoned Measure the similarity betweenimages by comparing the Hamming distance between theimage features of the query image and the image of thelibrary image In order to further improve the accuracy ofretrieval without affecting the real-time performance we canretrieve the image by the second level retrieval These stepsare described in detail as follows After one level retrievalwe choose the K images with the most similarity in thefirst-level retrieval result and then calculate the Euclideandistance between the original feature vector of the K imagesand the original feature vector of the query imageThe resultsobtained as the similarity measure of the images and outputthe retrieval result that has been ranked from the high andlow with the similarity distance
Although theHash algorithmmaps the high-dimensionalfeature vectors of the image into a hash-coded form theproblem of ldquodimensional disastersrdquo is solved and the retrievalefficiency is greatly accelerated However when the similaritycomparison is performed the Hamming distances of theimage features are simply compared using the results of theprimary search and occasionally undesirable results maystill appear on the search results If we want to increasethe accuracy of the search we must increase the hash codelength However excessively long codes will increase theamount of calculations increase the memory burden andreduce the real-time nature of retrieval failing to achieve
the goal of reducing the size of data In order to solve thisproblem keep the retrieval efficiency and further improvethe retrieval accuracy we propose a search strategy forsecondary retrieval the specific steps of which are as follows
Step 1 Through the first-level search in the Hamming spacethe similarity degree of the images is sorted and the top Ksorting images are selected
Step 2 For the 119896 images in Step 1 calculate the Euclideandistance one by one from its original image feature vector tothe image feature vector of the query image
Step 3 The Euclidean distance calculated in Step 2 is sortedThe smaller the calculated value is the higher similaritybetween images is and the similarity is sorted from high tolow and output as the final search result
In the second search it is necessary to pay attention tothe selection of the 119870 value although the larger the 119870 valueis the better the search effect is but accordingly the longerthe time is consumed Therefore it is necessary to combinevarious factors to select the appropriate119870 value
4 Experimental and Performance Analysis
In this section we thoroughly compare the proposedapproach with the improved deep learning hash retrievalmethods on several benchmark datasets Through a series ofexperiments the effectiveness and feasibility of the proposedalgorithm are verified
41 Database Two mostly used databases in the recent deeplearning hash works are taken into our evaluation Thetwo image libraries are derived from the CIFAR-10[11] coreexperimental image library dataset and theCaltech 256 imagelibrary dataset
CIFAR-10 dataset contains 10 object categories and eachclass consists of 6000 images resulting in a total of 60000images The dataset is split into training and test sets whichare averagely divided into 10 object classes The Caltech 256image library dataset contains 29780 color images which aregrouped intro 256 classes
First test the CIFAR-10 image dataset There are a totalof 50000 training samples which are used for training on thedeep learning networkThe remaining 10000 images are usedas test samples And then we randomly select 50 images fromdatabase as the query images For theHidden Image Retrievalalgorithm based on deep learning mentioned in this paperthe image pixel data is directly used as input while for otheralgorithms the 512-dimensional GIST feature is used as thefeature expression of the image Note quantization all imagesinto 32lowast32 sizes before experiment
For the Caltech 256 image a total of 256 classes areincluded and each class contains at least 70 images There-fore 70 images of each class a total of 17920 images arerandomly selected and are used as training images Theremaining images are used as test samples In addition all ofthe imagesrsquo size is set to 64lowast64 again when training
8 Security and Communication Networks
42 Evaluation Metrics We measure the performance ofcompared methods using Precision-recall and Average-Retrieval Precision (ARP) curves Precision is the ratio ofthe correct number of images m in the search result to thenumber k of all returned images The formula is as follows
precision = 119898119896 times 100 (12)
Recall is the ratio of the correct number of imagesm in thesearch results to the number g of images in the image libraryThe formula is as follows
recall = 119898119902 times 100 (13)
Assume that the search result of the query image 119894 is119861119894 and 119860 119894 means that the category is the same between thequery image and the return image then the accuracy rate forthe image query result 119875(119894) can be defined by the followingformula
119875 (119894) = |119860 (119894) cap 119861 (119894)||119861 (119894)| (14)
Average-Retrieval Precision (ARP) the average value ofall the images in the same class as the retrieval rate obtainedfrom the retrieval image is defined as follows
Here 119868119863119898 is the category index number of the image119898 isthe category index119873 is the number of imageswhose categoryis 119868119863119898 and 119894119889(119894) is the category index number of the queryimage
43 Performance Analysis In the proposed algorithm IDLHthe length of the hash sequence and the depth of the hiddenlayer in the deep learning network are two key parametersWhen the hash sequence length is small different featurevectors can easily be mapped into the same hash sequence sothe retrieval accuracy is low However if the hash sequence istoo long a large storage space is required and a long time isconsumed which reduces the real-time performance For thenumber of hidden layers the number of layers in the hiddenlayer is too small which is not conducive to learning strongimage features However if the depth of the hidden layer istoo large the difficulty of training is increased In order toverify the effectiveness and feasibility of our algorithm weconducted the following experiments
(1) Results on CIFAR-10 Dataset Figure 4(a) shows the searchresults of the Average-Retrieval Precision using our proposedalgorithm IDLH compared with the LSH algorithm [3] andother three deep learning algorithms the DH algorithm [21]the DeepBit algorithm [40] and the UH-BDNN algorithm[41] on CIFAR-10 dataset with 8 16 32 48 64 96 128and 256 bits Figure 4(b) shows the Precision-recall curveunder 48-bit encoding It can be seen that the algorithm hasa higher precision than the other hashing algorithms with
the same recall rate However the advantage is not obviousand the average accuracy is slightly higher than other hashalgorithms
In order to overcome the above defects we use deeplearning to perform hash mapping on image features per-form hash encoding of different bits on the same featurecalculate the Precision-recall of the search results under thecondition of different coded bits and determine the impactof the encoding length on the retrieval results
As Figure 5 shows with the increase in the number ofcoded bits the Precision-recall is continuously increasingWith the increase in the number of coded bits the imageis better expressed However after the number of coded bitsreaches 64 even if the number of coded bits increases theaverage accuracy rate increases relatively slowly Because theinformation of the tiny image is relatively simple when thenumber of encoded bits reaches 64 bits a relatively goodimage expression has been obtained and the performanceof the algorithm has basically stabilized At this time despiteincrease in the number of encoding bits it is not very helpfulto improve the accuracy rate
In addition we want to test the influence of the numberof hidden layers in the deep learning network on the retrievalresult by changing the number of hidden layers
Figure 6 shows the effect of deep learning networks onexperimental results in the case of different hidden layernumbers
It can be seen that deeper networks do not have muchimprovement in performance which is different from theexpectation that more hidden layers will help learn strongerimage features Since the image library data used in theexperiment is a tiny image library relatively good image char-acteristics can be learned using a deep learning network withfewer layers However if the image library is replaced witha more colorful image the deep neural network can acquiremore detailed image features and the deepening of thelearning network will greatly help the study of image features
(2) Results on Caltech 256 Image Data Set Figure 7(a)shows the results of the Average-Retrieval Precision resultswhen the number of coded bits is different Compared withthe black and white image library the proposed algorithmembodies the advantage of image feature learning and leadsthe Average-Retrieval Precision to other hash retrieval algo-rithms In Figure 7(b) we can also see that the algorithmproposed in this paper has a higher Precision-recall thanother algorithms under the same recall rate and it has bettersearch performance
As shown in Figure 8 as the number of coded bitsincreases the precision rate increases with the same recallrate This feature of the color image library is more pro-nounced than the black and white image library Becausethe color image contains more information more codingis needed to express it and the increase of encoding helpsto learn the features of the image The experimental resultsalso show that the deep learning network has learned moreexcellent image features
In Figure 9 the precision rate is significantly improvedby the increase in the number of hidden layers in the
Security and Communication Networks 9
IDLHLSHDHDeepBitUH-BDNN
16 32 48 64 96 128 2568Bits
0
01
02
03
04
05
06
07
08Av
erag
e Ret
rieva
l Pre
cisio
n
(a) Average-Retrieval Precision
IDLHLSHDHDeepBitUH-BDNN
01 02 03 04 05 06 07 08 09 100Recall
0
01
02
03
04
05
06
07
08
09
Prec
ision
(b) Precision-recall at 48 bits
Figure 4 Five kinds of algorithm retrieval performance comparison on CIFAR-10 dataset
8 bits16 bits32 bits48 bits64 bits128 bits
01 02 03 04 05 06 07 08 09 10Recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
Figure 5 Precision-recall curves with lengths
color image library Caltech 256 This is because theinformation contained in a more colorful image is morecomplex Adding a hidden layer can learn more detailsof the image and help improve the accuracy of thesearch
Next we tested the performance of secondary imageretrieval The value of k in the secondary search is 20 andthe experimental results are shown in Figure 10 As can
Level 1Level 2Level 3Level 4
01 02 03 04 05 06 07 08 09 1000Recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
Figure 6 Precision-recall curves with different code differenthidden layers
be seen from the results secondary retrieval can effectivelyimprove the retrieval accuracy when the number of codedbits is small However with the increase in the number ofencoding bits the results of the secondary search and theaccuracy of the primary search are not much different Thisis because the shorter the hash sequence is the easier thefeature vectors with different original features are mappedto the same hash code In order to make up for the errors
10 Security and Communication Networks
IDLHLSHDHDeepBitUH-BDNN
16 32 48 64 96 128 2568Bits
0
01
02
03
04
05
06
07
08
09
1Av
erag
e-Re
trie
val P
reci
sion
(a) Average-Retrieval Precision
IDLHLSHDHDeepBitUH-BDNN
01 02 03 04 05 06 07 08 09 1000Precision-recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
(b) Precision-recall at 48 bits
Figure 7 Five kinds of algorithm retrieval performance comparison on Caltech 256 set
8 bits16 bits32 bits48 bits64 bits128 bits
01 02 03 04 05 06 07 08 09 10Recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
Figure 8 Precision-recall curves with different code lengths
caused by the short hash code it is necessary to performsecondary searchWhen the number of encoding bits is smalla secondary retrieval method is used in IDLH and the searchaccuracy rate can be improved at the expense of a small searchspeed
Level 1Level 2Level 3Level 4
01 02 03 04 05 06 07 08 09 10Recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
Figure 9 Precision-recall curves with different hidden layers
5 Conclusion
With the rapid development of data storage and digital pro-cess more and more digital information is transformed andtransmitted over the Internet day by day which brings peoplea series of security problems as well as convenience [42] Theresearches on digital image security that is image encryptionimage data hiding and image authentication become moreimportant than ever [38 39] The most essential problem of
Security and Communication Networks 11
First retrievalSecond retrieval
16 32 48 64 96 128 2568Bits
07
08
09
1
Aver
age-
Retr
ieva
l Pre
cisio
n
Figure 10 Average-Retrieval Precision with first and secondretrieval
image recognition is to extract robust features The qualityof feature extraction is directly related to the effect of recog-nition so most of the previous work on image recognitionis spent on artificial design features [43] In recent yearsthe emergence of deep learning technology has changed thestatus of artificial design classification characteristics Deeplearning technology simulates the mechanism of humanvisual system information classification processing from themost primitive image pixels to lower edge features then tothe target components that are combined on the edge andfinally to the whole target depth learning can be combinedby layer by layer The high-level feature is the combination oflow level features From low level to high level features aremore and more abstract and show semantics more and moreFrom the underlying features to the combination of high-levelfeatures it is the depth of learning that is done by itself Itdoes not require manual intervention Compared with thecharacteristics of the artificial design this combination offeatures can be closer to the semantic expression
In terms of illegal image retrieval the traditional recog-nition method should establish a recognition model foreach type of recognition task In the actual application arecognition model needs a recognition server If there aremany identification tasks the cost is too high We used thedeep neural network to recognize the illegal image it onlyneeds to collect the samples of every kind of illegal imageand participate in the training of the deep neural networkFinally a multiclassification recognition model is trainedWhen classifying unknown samples deep neural networkaccounting calculates the probability that the image belongsto each class
We all know that in the image detection process theaccuracy and recall rate are mutually influential Ideally bothmust be high but in general the accuracy is high and the
recall rate is low the recall rate is high and the accuracy islow For image retrieval we need to improve the accuracyunder the condition of guaranteeing the recall rate Forimage disease surveillance and anti-illegal images we need toenhance the recall under the condition of ensuring accuracyTherefore in different application scenarios in order toachieve a balance between accuracy and recall perhaps somegame theory (such as Nash Equilibrium [44 45]) and penaltyfunction [46ndash48] can provide related optimization solutions
In this paper we proposed an improved deep-learning-hashing approach IDLH which optimized over two majorimage retrieval process
(a) In the feature extraction process the self-encodednetwork of the look-ahead type is trained by using unlabeledimage data and the expression of robust image features islearned This unlabeled learning method does not requireimage library labeling and reduces the requirements for theimage library At the same time it also takes advantage of thedeep learning networks strong learning ability and obtainsbetter image feature expression than ordinary algorithms
(b) On the index structure a secondary search is pro-posed which further increases the accuracy of the search atthe expense of very little retrieval time
Through experiments the algorithm proposed in thispaper is compared with other classic hashing algorithms onmultiple evaluation indicators Firstly we tested the learningnetworks of different code lengths and depths in order totest their effect on the retrieval system and then tested theperformance of the secondary search Through the above-mentioned series of experiments for different parameters theeffectiveness of the improved deep learning hash retrievalalgorithm proposed in this paper is verified and throughthe experimental data the good retrieval results are provedIn addition the proposed deep hashing training strategycan also be potentially applied to other hashing problemsinvolving data similarity computation
Data Availability
The data used to support the findings of this study areincluded within the article
Conflicts of Interest
The authors declare that they have no conflicts of interest
Acknowledgments
The work was funded by the National Natural ScienceFoundation of China (Grants nos 61206138 and 61373016)
References
[1] R Datta D Joshi J Li and J Z Wang ldquoImage retrieval ideasinfluences and trends of the new agerdquoACMComputing Surveysvol 40 no 2 article 5 2008
[2] G Shakhnarovich T Darrell and P Indyk Nearest-NeighborMethods in Learning andVisionTheory and PracticeMITPressCambridge MA USA 2006
12 Security and Communication Networks
[3] A Gionis P Indyk and R Motwani ldquoSimilarity search in highdimensions via hashingrdquo in 25th Int Conf pp 518ndash529 1999
[4] Z Pan J Lei Y Zhang and F L Wang ldquoAdaptive fractional-Pixel motion estimation skipped algorithm for efficient HEVCmotion estimationrdquoACMTransactions onMultimedia Comput-ing Communications and Applications (TOMM) vol 14 no 1pp 1ndash19 2018
[5] G-L Tian M Wang and L Song ldquoVariable selection in thehigh-dimensional continuous generalized linear model withcurrent status datardquo Journal of Applied Statistics vol 41 no 3pp 467ndash483 2014
[6] M Datar N Immorlica P Indyk and V S Mirrokni ldquoLocality-sensitive hashing scheme based on p-stable distributionsrdquo inProceedings of the 20th Annual Symposium on ComputationalGeometry (SCG rsquo04) pp 253ndash262 ACM June 2004
[7] B Kulis P Jain and K Grauman ldquoFast similarity search forlearned metricsrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 31 no 12 pp 2143ndash2157 2009
[8] M Raginsky and S Lazebnik ldquoLocality-sensitive binary codesfrom shift-invariant kernelsrdquo in Proceedings of the 23rd AnnualConference on Neural Information Processing Systems NIPS2009 pp 1509ndash1517 Canada December 2009
[9] L Qi X Zhang W Dou and Q Ni ldquoA distributed locality-sensitive hashing-based approach for cloud service recommen-dation from multi-source datardquo IEEE Journal on Selected Areasin Communications vol 35 no 11 pp 2616ndash2624 2017
[10] M A Carreira-Perpinan and R Raziperchikolaei ldquoHashingwith binary autoencodersrdquo in Proceedings of the IEEE Confer-ence on Computer Vision and Pattern Recognition CVPR 2015pp 557ndash566 USA June 2015
[11] J Wang S Kumar and S-F Chang ldquoSemi-supervised hashingfor large-scale searchrdquo IEEE Transactions on Pattern Analysisand Machine Intelligence vol 34 no 12 pp 2393ndash2406 2012
[12] Y Gong S Lazebnik and A Gordo ldquoIterative quantizationa Procrustean approach to learning binary codes for large-scale image retrievalrdquo in Proceedings of the IEEE Conference onComputer Vision and Pattern Recognition (CVPR rsquo11) pp 2916ndash2929 June 2011
[13] W Kong and W J Li ldquoIsotropic hashingrdquo NIPS vol 25 2012[14] M Norouzi and D J Fleet ldquoMinimal loss hashing for compact
binary codesrdquo in Proceedings of the 28th International Confer-ence on Machine Learning ICML 2011 pp 353ndash360 USA July2011
[15] J Wang W Liu A X Sun and Y-G Jiang ldquoLearning hashcodes with listwise supervisionrdquo in Proceedings of the 2013 14thIEEE International Conference on Computer Vision ICCV 2013pp 3032ndash3039 Australia December 2013
[16] G Lin C Shen Q Shi A Van Den Hengel and D Suter ldquoFastsupervised hashing with decision trees for high-dimensionaldatardquo in Proceedings of 27th IEEE Conference on ComputerVision and Pattern Recognition CVPRrsquo pp 1971ndash1978 USA2014
[17] Y Gong S Kumar H A Rowley and S Lazebnik ldquoLearningbinary codes for high-dimensional data using bilinear projec-tionsrdquo in Proceedings of the 26th IEEE Conference on ComputerVision and Pattern Recognition CVPR 2013 pp 484ndash491 USAJune 2013
[18] W Liu J Wang Y Mu and S Kumar ldquoCompact hyperplanehashing with bilinear functionsrdquo in The 29th InternationalConference on Machine Learning (ICML12) pp 467ndash474 2012
[19] Y Weiss A Torralba and R Fergus ldquoSpectral hashingrdquo inProceedings of the 22nd Annual Conference on Neural Informa-tion Processing Systems (NIPS rsquo08) pp 1753ndash1760 VancouverCanada December 2008
[20] W Liu J Wang S Kumar and S F Chang ldquoHashing withgraphsrdquo inThe 28th international conference on machine learn-ing (ICML11) 2011
[21] F Shen X Zhou Y Yang J Song H T Shen and D Tao ldquoA fastoptimization method for general binary code learningrdquo IEEETransactions on Image Processing vol 25 no 12 pp 5610ndash56212016
[22] F Shen W Liu S Zhang Y Yang and H T Shen ldquoLearningbinary codes for maximum inner product searchrdquo in Proceed-ings of the 15th IEEE International Conference on ComputerVision ICCV 2015 pp 4148ndash4156 Chile December 2015
[23] A Krizhevsky I Sutskever andG EHinton ldquoImagenet classifi-cation with deep convolutional neural networksrdquo in Proceedingsof the 26th Annual Conference on Neural Information ProcessingSystems (NIPS rsquo12) pp 1097ndash1105 Lake Tahoe Nev USADecember 2012
[24] G E Hinton and R R Salakhutdinov ldquoReducing the dimen-sionality of data with neural networksrdquo The American Associa-tion for the Advancement of Science Science vol 313 no 5786pp 504ndash507 2006
[25] A Torralba R Fergus and Y Weiss ldquoSmall codes and largeimage databases for recognitionrdquo in Proceedings of the IEEEComputer Society Conference on Computer Vision and PatternRecognition (CVPR rsquo08) pp 1ndash8 2008
[26] R Salakhutdinov andG Hinton ldquoLearning a nonlinear embed-ding by preserving class neighbourhood structurerdquo Journal ofMachine Learning Research vol 2 pp 412ndash419 2007
[27] V E Liong J Lu GWang P Moulin and J Zhou ldquoDeep hash-ing for compact binary codes learningrdquo in Proceedings of theIEEE Conference on Computer Vision and Pattern RecognitionCVPR 2015 pp 2475ndash2483 USA June 2015
[28] Y Gong S Lazebnik A Gordo and F Perronnin ldquoIterativequantization A procrustean approach to learning binary codesfor large-scale image retrievalrdquo IEEE Transactions on PatternAnalysis andMachine Intelligence vol 35 no 12 pp 2916ndash29292013
[29] W Liu J Wang R Ji Y-G Jiang and S-F Chang ldquoSupervisedhashing with kernelsrdquo in Proceedings of the IEEE Conference onComputer Vision and Pattern Recognition (CVPR rsquo12) pp 2074ndash2081 Providence RI USA June 2012
[30] J Masci A Bronstein M Bronstein and P SprechmannldquoSparse similarity-preserving hashingrdquo in Int Conf LearnRepresent pp 1ndash13 2014
[31] B Kulis and K Grauman ldquoKernelized locality-sensitive hash-ingrdquo IEEE Transactions on Pattern Analysis and Machine Intel-ligence vol 34 no 6 pp 1092ndash1104 2012
[32] F Zhao YHuang LWang and T Tan ldquoDeep semantic rankingbased hashing for multi-label image retrievalrdquo in Proceedings ofIEEE Conference on Computer Vision and Pattern RecognitionCVPR 2015 pp 1556ndash1564 June 2015
[33] G Cheng C Yang X Yao L Guo and J Han ldquoWhenDeep Learning Meets Metric Learning Remote Sensing ImageScene Classification via Learning Discriminative CNNsrdquo IEEETransactions on Geoscience and Remote Sensing pp 1ndash11
[34] J He J Feng X Liu et al ldquoMobile product search with Bag ofHash Bits and boundary rerankingrdquo in Proceedings of the 2012IEEE Conference on Computer Vision and Pattern RecognitionCVPR 2012 pp 3005ndash3012 USA June 2012
Security and Communication Networks 13
[35] F Shen Y Mu Y Yang et al ldquoClassification by retrievalBinarizing data and classifiersrdquo in Proceedings of the 40thInternational ACM SIGIR Conference on Research and Develop-ment in Information Retrieval SIGIR 2017 pp 595ndash604 JapanAugust 2017
[36] P Li S Zhao andR Zhang ldquoA cluster analysis selection strategyfor supersaturated designsrdquo Computational Statistics amp DataAnalysis vol 54 no 6 pp 1605ndash1612 2010
[37] A Pradeep S Mridula and P Mohanan ldquoHigh securityidentity tags using spiral resonatorsrdquo Cmc-Computers Materialsamp Continua vol 52 no 3 pp 187ndash196 2016
[38] Y Cao Z Zhou X Sun and C Gao ldquoCoverless informationhiding based on the molecular structure images of materialrdquoComputers Materials and Continua vol 54 no 2 pp 197ndash2072018
[39] Y LiuH Peng and JWang ldquoVerifiable diversity ranking searchover encrypted outsourced datardquo Cmc-Computers Materials ampContinua vol 55 no 1 pp 037ndash057 2018
[40] K Lin J Lu C-S Chen and J Zhou ldquoLearning compactbinary descriptors with unsupervised deep neural networksrdquo inProceedings of the 2016 IEEEConference onComputer Vision andPattern Recognition CVPR 2016 pp 1183ndash1192 USA July 2016
[41] T Do A Doan and N Cheung ldquoLearning to Hash with BinaryDeep Neural Networkrdquo in Computer Vision ndash ECCV 2016vol 9909 of Lecture Notes in Computer Science pp 219ndash234Springer International Publishing Cham 2016
[42] Rui Zhang Di Xiao and Yanting Chang ldquoA Novel ImageAuthentication with Tamper Localization and Self-Recovery inEncrypted Domain Based on Compressive Sensingrdquo Securityand Communication Networks vol 2018 Article ID 1591206 15pages 2018
[43] Xia ShuangKui and JianbinWu ldquoAModification-Free Steganog-raphy Method Based on Image Information Entropyrdquo Securityand Communication Networks vol 2018 Article ID 6256872 8pages 2018
[44] J Zhang B Qu and N Xiu ldquoSome projection-like methods forthe generalized Nash equilibriardquo Computational Optimizationand Applications vol 45 no 1 pp 89ndash109 2010
[45] Biao Qu and Jing Zhao ldquoMethods for Solving Generalized NashEquilibriumrdquo Journal of Applied Mathematics vol 2013 ArticleID 762165 6 pages 2013
[46] CWang CMa and J Zhou ldquoA new class of exact penalty func-tions and penalty algorithmsrdquo Journal of Global Optimizationvol 58 no 1 pp 51ndash73 2014
[47] Y Wang X Sun and F Meng ldquoOn the conditional andpartial trade credit policywith capital constraints A StackelbergModelrdquo Applied Mathematical Modelling vol 40 no 1 pp 1ndash182016
[48] S Lian and Y Duan ldquoSmoothing of the lower-order exactpenalty function for inequality constrained optimizationrdquo Jour-nal of Inequalities and Applications Paper No 185 12 pages2016
Chemical EngineeringInternational Journal of Antennas and
Propagation
International Journal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Navigation and Observation
International Journal of
Hindawi
wwwhindawicom Volume 2018
Advances in
Multimedia
Submit your manuscripts atwwwhindawicom
Security and Communication Networks 5
Hash Layer 2
Hamming Retrieval
Query Result
Weights
Codes
Training Dataset
Feature expression
function
Image Preprocess Layer Feature
expression function
Hash Layer 1
Binary Weights
Binary Codes
Feature migration
Query Image
Figure 1 Deep learning-hash retrieval framework IDLH consists of three main components (preprocessing layer deep neural networklayer and hash layer) The object of the first layer is simply scaled sample-by-sample mean-value reduction and whitening In the secondcomponent we develop a deep neural network to obtain the feature expression function of the image And the classifier weights and featurebinary codes are simultaneously learned in the last component-hash layer
layer can be obtained and the output of the hidden layer canbe used as the input of the second self-encoder to continuetraining the second self-encoder Then the output ℎ(119896)2 of thesecond hidden layer of the self-encoder can be obtained Theoutputℎ(119896)1 of the first hidden layer of the self-encoder is calleda first-order feature and the output ℎ(119896)2 of the second hiddenlayer of the self-encoder is called a second-order feature Inorder to classify the two-order feature ℎ(119896)2 can be used as theinput of Softmax regression
Figure 2 shows the flowchart of the proposed methodAnd there are mainly three processes (supervised pretrain-ing rough image retrieval and accurate image retrieval)The object of the first process is to transform the high-dimensional feature vector into a low-dimensional compacttwo value codes through hash function In the secondprocedure we pick out M candidate images by calculatingHamming distance In the third process we calculate theEuclidean distance between the candidate image and theimage to be retrieved and accurately extract K images fromthe M candidate images
Figure 3 shows a block diagram of a self-encoding neuralnetwork The stacking self-encoding network contains 3hidden layers (feature layers) The input layer inputs theoriginal data i into the first layer of the feature layer theoutput result of the former layer serves as the input of thenext layer and the output of the third layer serves as thefeature expression of the image In our method it is also usedas the input of the hash classifier and it is possible to usethe characteristics of the STD neural network to classify thefeatures
By using the matrix representation of the binary codesvectors and the output of the 3th layer of the network we usethe gradient descent method to solve the neural network
For a single sample (x y) the cost function is as shown in
(12 10038171003817100381710038171003817ℎ119882119887 (119909(119894)) minus 119910(119894)100381710038171003817100381710038172)]
+ 1205822119899119897minus1sum119897=1
119904119897sum119894=1
119904119897+1sum119895=1
(119882(119897)119895119894 )2
(6)
The first term 119869(119882 119887) represents the mean variance termThe second term aims to prevent the data from overfitting byreducing the magnitude of the weight 120582 is a weight attenua-tion parameter It is used to balance the relative importance ofmean square deviation terms and weight attenuation termsOur purpose is to minimize the quantization loss 119869(119882 119887)between the learned binary values and the real values of aninput image according to parameters119882 and 11988733 Hash Algorithm Retrieval The image retrieval methodbased on hash algorithmmaps the high-dimensional contentfeatures of images into Hamming space (binary space) andgenerates a low-dimensional hash sequence to represent apicture This method reduces the requirement of computermemory space for image retrieval system improves theretrieval speed and better adapts to the requirements of massimage retrieval
Inspired by [6 8] we use a set of hash functions to hashdata into different buckets After we do some hash mappingon the original image feature data we hope that the originaltwo adjacent feature data can be hash into the same bucketwith the same bucket number And then after hash mappingof all the data in the original feature set we can get a hashtable These original feature data sets are scattered into hashtable buckets and the data belonging to the same bucket isprobably adjacent to the original data However there is also
6 Security and Communication Networks
StartTwo valued by the
Sigmod function in Hash layer
Building hash pool and hash bucket
Calculation of Hamming distance
Select the top m candidate images
Calculation of Euclidean distance
Select the top kimages from m
candidate imagesas retrieval result
End
Supervised pre-training
Rough image retrieval
If itgt05Yes
1
No
0
If it less than a threshold
Yes
No
Accurate image retrieval
The featureof the image
to be retrieved
Self-learning network
The featureof the
training images
Binary processing
Binary codes for trainingimage
Binary codes for retrieving
image
Figure 2 Deep learning-hash retrieval flowchart IDLHmainly includes three processes (supervised pretraining rough image retrieval andaccurate image retrieval)
+1+1 +1
+1
Layer 0Input layer
Layer 1 Layer 2 Layer 3 Hash Layer
x1
x2
x3
x4
x5
ℎ(1)1
ℎ(1)2
ℎ(1)3
ℎ(1)4
ℎ(2)1
ℎ(2)2
ℎ(2)3
ℎ(2)4
ℎ(3)1
ℎ(3)2
ℎ(3)3
Output ℎ(3)k
from layer 3 is usedas binary code with bk = f(ℎ(3)
k)
Figure 3 Self-learning network based on stack self-encoding network The neurons labeled xi is the input of the neural network and ldquo+ 1rdquoare the offset nodes (intercept entries) of the neural network The layer 0 is the input layer of neural network and layer 3 is the output layerof neural network The middle layers of layer 0 layer to layer 3 are the hidden layers of neural network
a small probability in events that is the nonadjacent data ishash to the same barrel Set the hash function as the follows
Here119908119896 is the projection vector and 119887119896 is the correspond-ing intercept The code value generated by formula (7) isminus1 1 and we use the following formula to convert it intotwo value codes
119910119896 = 12 (1 + ℎ119896 (119909)) (8)
Given a sample point isin 119877119863 we can compute a K-bitbinary code 119910 for 119909 with formula (9) The hash functionperforms the mapping as ℎ119896 119877119863 997888rarr 119861
Here 119883 = 119909119899119873119899=1 isin 119877119863times119873 is the feature data matrix withpoints as columns Such a binary encoding process can also beviewed as mapping the original data point to a binary valuedspace namely Hamming space
34 SimilarityMeasure After obtaining the binary hash codeof the image it is necessary tomeasure similarity between theretrieved image and the library image in the Hamming spaceThe smaller the Hamming distance is the closer distancebetween the two data is and the degree of similarity is higherotherwise the two data similarity is lower
Here oplus is an XOR operation The two sets of 119910119894 and 119910119895represent the hash code of the search image feature and theimage library is mapped through the hash function The newimage features learned by the stack self-encoding network aregenerated by the hash function The storage structure of theimage feature vectors is shown in Table 3
As can be seen from Table 3 the hash code of the imageis related to the image ID and the image name one by one Inthe process of searching the image feature vector is obtainedthrough deep learning by a hash function the original data ismapped into a newdata space and a corresponding hash codeis obtained The hash code is used to calculate the Hammingdistance in the Hamming space as a measure of similaritybetween images Finally the storage structure of the imagefeature vector is used to find the corresponding image ID ofthe hash code and the output retrieval result is output to theuser
35 Image Secondary Search In the first-level search phasethe features learned from the deep learning network aremapped into the Hamming space using the hash function Inthe similarity measurement phase the traditional Euclideandistance is abandoned Measure the similarity betweenimages by comparing the Hamming distance between theimage features of the query image and the image of thelibrary image In order to further improve the accuracy ofretrieval without affecting the real-time performance we canretrieve the image by the second level retrieval These stepsare described in detail as follows After one level retrievalwe choose the K images with the most similarity in thefirst-level retrieval result and then calculate the Euclideandistance between the original feature vector of the K imagesand the original feature vector of the query imageThe resultsobtained as the similarity measure of the images and outputthe retrieval result that has been ranked from the high andlow with the similarity distance
Although theHash algorithmmaps the high-dimensionalfeature vectors of the image into a hash-coded form theproblem of ldquodimensional disastersrdquo is solved and the retrievalefficiency is greatly accelerated However when the similaritycomparison is performed the Hamming distances of theimage features are simply compared using the results of theprimary search and occasionally undesirable results maystill appear on the search results If we want to increasethe accuracy of the search we must increase the hash codelength However excessively long codes will increase theamount of calculations increase the memory burden andreduce the real-time nature of retrieval failing to achieve
the goal of reducing the size of data In order to solve thisproblem keep the retrieval efficiency and further improvethe retrieval accuracy we propose a search strategy forsecondary retrieval the specific steps of which are as follows
Step 1 Through the first-level search in the Hamming spacethe similarity degree of the images is sorted and the top Ksorting images are selected
Step 2 For the 119896 images in Step 1 calculate the Euclideandistance one by one from its original image feature vector tothe image feature vector of the query image
Step 3 The Euclidean distance calculated in Step 2 is sortedThe smaller the calculated value is the higher similaritybetween images is and the similarity is sorted from high tolow and output as the final search result
In the second search it is necessary to pay attention tothe selection of the 119870 value although the larger the 119870 valueis the better the search effect is but accordingly the longerthe time is consumed Therefore it is necessary to combinevarious factors to select the appropriate119870 value
4 Experimental and Performance Analysis
In this section we thoroughly compare the proposedapproach with the improved deep learning hash retrievalmethods on several benchmark datasets Through a series ofexperiments the effectiveness and feasibility of the proposedalgorithm are verified
41 Database Two mostly used databases in the recent deeplearning hash works are taken into our evaluation Thetwo image libraries are derived from the CIFAR-10[11] coreexperimental image library dataset and theCaltech 256 imagelibrary dataset
CIFAR-10 dataset contains 10 object categories and eachclass consists of 6000 images resulting in a total of 60000images The dataset is split into training and test sets whichare averagely divided into 10 object classes The Caltech 256image library dataset contains 29780 color images which aregrouped intro 256 classes
First test the CIFAR-10 image dataset There are a totalof 50000 training samples which are used for training on thedeep learning networkThe remaining 10000 images are usedas test samples And then we randomly select 50 images fromdatabase as the query images For theHidden Image Retrievalalgorithm based on deep learning mentioned in this paperthe image pixel data is directly used as input while for otheralgorithms the 512-dimensional GIST feature is used as thefeature expression of the image Note quantization all imagesinto 32lowast32 sizes before experiment
For the Caltech 256 image a total of 256 classes areincluded and each class contains at least 70 images There-fore 70 images of each class a total of 17920 images arerandomly selected and are used as training images Theremaining images are used as test samples In addition all ofthe imagesrsquo size is set to 64lowast64 again when training
8 Security and Communication Networks
42 Evaluation Metrics We measure the performance ofcompared methods using Precision-recall and Average-Retrieval Precision (ARP) curves Precision is the ratio ofthe correct number of images m in the search result to thenumber k of all returned images The formula is as follows
precision = 119898119896 times 100 (12)
Recall is the ratio of the correct number of imagesm in thesearch results to the number g of images in the image libraryThe formula is as follows
recall = 119898119902 times 100 (13)
Assume that the search result of the query image 119894 is119861119894 and 119860 119894 means that the category is the same between thequery image and the return image then the accuracy rate forthe image query result 119875(119894) can be defined by the followingformula
119875 (119894) = |119860 (119894) cap 119861 (119894)||119861 (119894)| (14)
Average-Retrieval Precision (ARP) the average value ofall the images in the same class as the retrieval rate obtainedfrom the retrieval image is defined as follows
Here 119868119863119898 is the category index number of the image119898 isthe category index119873 is the number of imageswhose categoryis 119868119863119898 and 119894119889(119894) is the category index number of the queryimage
43 Performance Analysis In the proposed algorithm IDLHthe length of the hash sequence and the depth of the hiddenlayer in the deep learning network are two key parametersWhen the hash sequence length is small different featurevectors can easily be mapped into the same hash sequence sothe retrieval accuracy is low However if the hash sequence istoo long a large storage space is required and a long time isconsumed which reduces the real-time performance For thenumber of hidden layers the number of layers in the hiddenlayer is too small which is not conducive to learning strongimage features However if the depth of the hidden layer istoo large the difficulty of training is increased In order toverify the effectiveness and feasibility of our algorithm weconducted the following experiments
(1) Results on CIFAR-10 Dataset Figure 4(a) shows the searchresults of the Average-Retrieval Precision using our proposedalgorithm IDLH compared with the LSH algorithm [3] andother three deep learning algorithms the DH algorithm [21]the DeepBit algorithm [40] and the UH-BDNN algorithm[41] on CIFAR-10 dataset with 8 16 32 48 64 96 128and 256 bits Figure 4(b) shows the Precision-recall curveunder 48-bit encoding It can be seen that the algorithm hasa higher precision than the other hashing algorithms with
the same recall rate However the advantage is not obviousand the average accuracy is slightly higher than other hashalgorithms
In order to overcome the above defects we use deeplearning to perform hash mapping on image features per-form hash encoding of different bits on the same featurecalculate the Precision-recall of the search results under thecondition of different coded bits and determine the impactof the encoding length on the retrieval results
As Figure 5 shows with the increase in the number ofcoded bits the Precision-recall is continuously increasingWith the increase in the number of coded bits the imageis better expressed However after the number of coded bitsreaches 64 even if the number of coded bits increases theaverage accuracy rate increases relatively slowly Because theinformation of the tiny image is relatively simple when thenumber of encoded bits reaches 64 bits a relatively goodimage expression has been obtained and the performanceof the algorithm has basically stabilized At this time despiteincrease in the number of encoding bits it is not very helpfulto improve the accuracy rate
In addition we want to test the influence of the numberof hidden layers in the deep learning network on the retrievalresult by changing the number of hidden layers
Figure 6 shows the effect of deep learning networks onexperimental results in the case of different hidden layernumbers
It can be seen that deeper networks do not have muchimprovement in performance which is different from theexpectation that more hidden layers will help learn strongerimage features Since the image library data used in theexperiment is a tiny image library relatively good image char-acteristics can be learned using a deep learning network withfewer layers However if the image library is replaced witha more colorful image the deep neural network can acquiremore detailed image features and the deepening of thelearning network will greatly help the study of image features
(2) Results on Caltech 256 Image Data Set Figure 7(a)shows the results of the Average-Retrieval Precision resultswhen the number of coded bits is different Compared withthe black and white image library the proposed algorithmembodies the advantage of image feature learning and leadsthe Average-Retrieval Precision to other hash retrieval algo-rithms In Figure 7(b) we can also see that the algorithmproposed in this paper has a higher Precision-recall thanother algorithms under the same recall rate and it has bettersearch performance
As shown in Figure 8 as the number of coded bitsincreases the precision rate increases with the same recallrate This feature of the color image library is more pro-nounced than the black and white image library Becausethe color image contains more information more codingis needed to express it and the increase of encoding helpsto learn the features of the image The experimental resultsalso show that the deep learning network has learned moreexcellent image features
In Figure 9 the precision rate is significantly improvedby the increase in the number of hidden layers in the
Security and Communication Networks 9
IDLHLSHDHDeepBitUH-BDNN
16 32 48 64 96 128 2568Bits
0
01
02
03
04
05
06
07
08Av
erag
e Ret
rieva
l Pre
cisio
n
(a) Average-Retrieval Precision
IDLHLSHDHDeepBitUH-BDNN
01 02 03 04 05 06 07 08 09 100Recall
0
01
02
03
04
05
06
07
08
09
Prec
ision
(b) Precision-recall at 48 bits
Figure 4 Five kinds of algorithm retrieval performance comparison on CIFAR-10 dataset
8 bits16 bits32 bits48 bits64 bits128 bits
01 02 03 04 05 06 07 08 09 10Recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
Figure 5 Precision-recall curves with lengths
color image library Caltech 256 This is because theinformation contained in a more colorful image is morecomplex Adding a hidden layer can learn more detailsof the image and help improve the accuracy of thesearch
Next we tested the performance of secondary imageretrieval The value of k in the secondary search is 20 andthe experimental results are shown in Figure 10 As can
Level 1Level 2Level 3Level 4
01 02 03 04 05 06 07 08 09 1000Recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
Figure 6 Precision-recall curves with different code differenthidden layers
be seen from the results secondary retrieval can effectivelyimprove the retrieval accuracy when the number of codedbits is small However with the increase in the number ofencoding bits the results of the secondary search and theaccuracy of the primary search are not much different Thisis because the shorter the hash sequence is the easier thefeature vectors with different original features are mappedto the same hash code In order to make up for the errors
10 Security and Communication Networks
IDLHLSHDHDeepBitUH-BDNN
16 32 48 64 96 128 2568Bits
0
01
02
03
04
05
06
07
08
09
1Av
erag
e-Re
trie
val P
reci
sion
(a) Average-Retrieval Precision
IDLHLSHDHDeepBitUH-BDNN
01 02 03 04 05 06 07 08 09 1000Precision-recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
(b) Precision-recall at 48 bits
Figure 7 Five kinds of algorithm retrieval performance comparison on Caltech 256 set
8 bits16 bits32 bits48 bits64 bits128 bits
01 02 03 04 05 06 07 08 09 10Recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
Figure 8 Precision-recall curves with different code lengths
caused by the short hash code it is necessary to performsecondary searchWhen the number of encoding bits is smalla secondary retrieval method is used in IDLH and the searchaccuracy rate can be improved at the expense of a small searchspeed
Level 1Level 2Level 3Level 4
01 02 03 04 05 06 07 08 09 10Recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
Figure 9 Precision-recall curves with different hidden layers
5 Conclusion
With the rapid development of data storage and digital pro-cess more and more digital information is transformed andtransmitted over the Internet day by day which brings peoplea series of security problems as well as convenience [42] Theresearches on digital image security that is image encryptionimage data hiding and image authentication become moreimportant than ever [38 39] The most essential problem of
Security and Communication Networks 11
First retrievalSecond retrieval
16 32 48 64 96 128 2568Bits
07
08
09
1
Aver
age-
Retr
ieva
l Pre
cisio
n
Figure 10 Average-Retrieval Precision with first and secondretrieval
image recognition is to extract robust features The qualityof feature extraction is directly related to the effect of recog-nition so most of the previous work on image recognitionis spent on artificial design features [43] In recent yearsthe emergence of deep learning technology has changed thestatus of artificial design classification characteristics Deeplearning technology simulates the mechanism of humanvisual system information classification processing from themost primitive image pixels to lower edge features then tothe target components that are combined on the edge andfinally to the whole target depth learning can be combinedby layer by layer The high-level feature is the combination oflow level features From low level to high level features aremore and more abstract and show semantics more and moreFrom the underlying features to the combination of high-levelfeatures it is the depth of learning that is done by itself Itdoes not require manual intervention Compared with thecharacteristics of the artificial design this combination offeatures can be closer to the semantic expression
In terms of illegal image retrieval the traditional recog-nition method should establish a recognition model foreach type of recognition task In the actual application arecognition model needs a recognition server If there aremany identification tasks the cost is too high We used thedeep neural network to recognize the illegal image it onlyneeds to collect the samples of every kind of illegal imageand participate in the training of the deep neural networkFinally a multiclassification recognition model is trainedWhen classifying unknown samples deep neural networkaccounting calculates the probability that the image belongsto each class
We all know that in the image detection process theaccuracy and recall rate are mutually influential Ideally bothmust be high but in general the accuracy is high and the
recall rate is low the recall rate is high and the accuracy islow For image retrieval we need to improve the accuracyunder the condition of guaranteeing the recall rate Forimage disease surveillance and anti-illegal images we need toenhance the recall under the condition of ensuring accuracyTherefore in different application scenarios in order toachieve a balance between accuracy and recall perhaps somegame theory (such as Nash Equilibrium [44 45]) and penaltyfunction [46ndash48] can provide related optimization solutions
In this paper we proposed an improved deep-learning-hashing approach IDLH which optimized over two majorimage retrieval process
(a) In the feature extraction process the self-encodednetwork of the look-ahead type is trained by using unlabeledimage data and the expression of robust image features islearned This unlabeled learning method does not requireimage library labeling and reduces the requirements for theimage library At the same time it also takes advantage of thedeep learning networks strong learning ability and obtainsbetter image feature expression than ordinary algorithms
(b) On the index structure a secondary search is pro-posed which further increases the accuracy of the search atthe expense of very little retrieval time
Through experiments the algorithm proposed in thispaper is compared with other classic hashing algorithms onmultiple evaluation indicators Firstly we tested the learningnetworks of different code lengths and depths in order totest their effect on the retrieval system and then tested theperformance of the secondary search Through the above-mentioned series of experiments for different parameters theeffectiveness of the improved deep learning hash retrievalalgorithm proposed in this paper is verified and throughthe experimental data the good retrieval results are provedIn addition the proposed deep hashing training strategycan also be potentially applied to other hashing problemsinvolving data similarity computation
Data Availability
The data used to support the findings of this study areincluded within the article
Conflicts of Interest
The authors declare that they have no conflicts of interest
Acknowledgments
The work was funded by the National Natural ScienceFoundation of China (Grants nos 61206138 and 61373016)
References
[1] R Datta D Joshi J Li and J Z Wang ldquoImage retrieval ideasinfluences and trends of the new agerdquoACMComputing Surveysvol 40 no 2 article 5 2008
[2] G Shakhnarovich T Darrell and P Indyk Nearest-NeighborMethods in Learning andVisionTheory and PracticeMITPressCambridge MA USA 2006
12 Security and Communication Networks
[3] A Gionis P Indyk and R Motwani ldquoSimilarity search in highdimensions via hashingrdquo in 25th Int Conf pp 518ndash529 1999
[4] Z Pan J Lei Y Zhang and F L Wang ldquoAdaptive fractional-Pixel motion estimation skipped algorithm for efficient HEVCmotion estimationrdquoACMTransactions onMultimedia Comput-ing Communications and Applications (TOMM) vol 14 no 1pp 1ndash19 2018
[5] G-L Tian M Wang and L Song ldquoVariable selection in thehigh-dimensional continuous generalized linear model withcurrent status datardquo Journal of Applied Statistics vol 41 no 3pp 467ndash483 2014
[6] M Datar N Immorlica P Indyk and V S Mirrokni ldquoLocality-sensitive hashing scheme based on p-stable distributionsrdquo inProceedings of the 20th Annual Symposium on ComputationalGeometry (SCG rsquo04) pp 253ndash262 ACM June 2004
[7] B Kulis P Jain and K Grauman ldquoFast similarity search forlearned metricsrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 31 no 12 pp 2143ndash2157 2009
[8] M Raginsky and S Lazebnik ldquoLocality-sensitive binary codesfrom shift-invariant kernelsrdquo in Proceedings of the 23rd AnnualConference on Neural Information Processing Systems NIPS2009 pp 1509ndash1517 Canada December 2009
[9] L Qi X Zhang W Dou and Q Ni ldquoA distributed locality-sensitive hashing-based approach for cloud service recommen-dation from multi-source datardquo IEEE Journal on Selected Areasin Communications vol 35 no 11 pp 2616ndash2624 2017
[10] M A Carreira-Perpinan and R Raziperchikolaei ldquoHashingwith binary autoencodersrdquo in Proceedings of the IEEE Confer-ence on Computer Vision and Pattern Recognition CVPR 2015pp 557ndash566 USA June 2015
[11] J Wang S Kumar and S-F Chang ldquoSemi-supervised hashingfor large-scale searchrdquo IEEE Transactions on Pattern Analysisand Machine Intelligence vol 34 no 12 pp 2393ndash2406 2012
[12] Y Gong S Lazebnik and A Gordo ldquoIterative quantizationa Procrustean approach to learning binary codes for large-scale image retrievalrdquo in Proceedings of the IEEE Conference onComputer Vision and Pattern Recognition (CVPR rsquo11) pp 2916ndash2929 June 2011
[13] W Kong and W J Li ldquoIsotropic hashingrdquo NIPS vol 25 2012[14] M Norouzi and D J Fleet ldquoMinimal loss hashing for compact
binary codesrdquo in Proceedings of the 28th International Confer-ence on Machine Learning ICML 2011 pp 353ndash360 USA July2011
[15] J Wang W Liu A X Sun and Y-G Jiang ldquoLearning hashcodes with listwise supervisionrdquo in Proceedings of the 2013 14thIEEE International Conference on Computer Vision ICCV 2013pp 3032ndash3039 Australia December 2013
[16] G Lin C Shen Q Shi A Van Den Hengel and D Suter ldquoFastsupervised hashing with decision trees for high-dimensionaldatardquo in Proceedings of 27th IEEE Conference on ComputerVision and Pattern Recognition CVPRrsquo pp 1971ndash1978 USA2014
[17] Y Gong S Kumar H A Rowley and S Lazebnik ldquoLearningbinary codes for high-dimensional data using bilinear projec-tionsrdquo in Proceedings of the 26th IEEE Conference on ComputerVision and Pattern Recognition CVPR 2013 pp 484ndash491 USAJune 2013
[18] W Liu J Wang Y Mu and S Kumar ldquoCompact hyperplanehashing with bilinear functionsrdquo in The 29th InternationalConference on Machine Learning (ICML12) pp 467ndash474 2012
[19] Y Weiss A Torralba and R Fergus ldquoSpectral hashingrdquo inProceedings of the 22nd Annual Conference on Neural Informa-tion Processing Systems (NIPS rsquo08) pp 1753ndash1760 VancouverCanada December 2008
[20] W Liu J Wang S Kumar and S F Chang ldquoHashing withgraphsrdquo inThe 28th international conference on machine learn-ing (ICML11) 2011
[21] F Shen X Zhou Y Yang J Song H T Shen and D Tao ldquoA fastoptimization method for general binary code learningrdquo IEEETransactions on Image Processing vol 25 no 12 pp 5610ndash56212016
[22] F Shen W Liu S Zhang Y Yang and H T Shen ldquoLearningbinary codes for maximum inner product searchrdquo in Proceed-ings of the 15th IEEE International Conference on ComputerVision ICCV 2015 pp 4148ndash4156 Chile December 2015
[23] A Krizhevsky I Sutskever andG EHinton ldquoImagenet classifi-cation with deep convolutional neural networksrdquo in Proceedingsof the 26th Annual Conference on Neural Information ProcessingSystems (NIPS rsquo12) pp 1097ndash1105 Lake Tahoe Nev USADecember 2012
[24] G E Hinton and R R Salakhutdinov ldquoReducing the dimen-sionality of data with neural networksrdquo The American Associa-tion for the Advancement of Science Science vol 313 no 5786pp 504ndash507 2006
[25] A Torralba R Fergus and Y Weiss ldquoSmall codes and largeimage databases for recognitionrdquo in Proceedings of the IEEEComputer Society Conference on Computer Vision and PatternRecognition (CVPR rsquo08) pp 1ndash8 2008
[26] R Salakhutdinov andG Hinton ldquoLearning a nonlinear embed-ding by preserving class neighbourhood structurerdquo Journal ofMachine Learning Research vol 2 pp 412ndash419 2007
[27] V E Liong J Lu GWang P Moulin and J Zhou ldquoDeep hash-ing for compact binary codes learningrdquo in Proceedings of theIEEE Conference on Computer Vision and Pattern RecognitionCVPR 2015 pp 2475ndash2483 USA June 2015
[28] Y Gong S Lazebnik A Gordo and F Perronnin ldquoIterativequantization A procrustean approach to learning binary codesfor large-scale image retrievalrdquo IEEE Transactions on PatternAnalysis andMachine Intelligence vol 35 no 12 pp 2916ndash29292013
[29] W Liu J Wang R Ji Y-G Jiang and S-F Chang ldquoSupervisedhashing with kernelsrdquo in Proceedings of the IEEE Conference onComputer Vision and Pattern Recognition (CVPR rsquo12) pp 2074ndash2081 Providence RI USA June 2012
[30] J Masci A Bronstein M Bronstein and P SprechmannldquoSparse similarity-preserving hashingrdquo in Int Conf LearnRepresent pp 1ndash13 2014
[31] B Kulis and K Grauman ldquoKernelized locality-sensitive hash-ingrdquo IEEE Transactions on Pattern Analysis and Machine Intel-ligence vol 34 no 6 pp 1092ndash1104 2012
[32] F Zhao YHuang LWang and T Tan ldquoDeep semantic rankingbased hashing for multi-label image retrievalrdquo in Proceedings ofIEEE Conference on Computer Vision and Pattern RecognitionCVPR 2015 pp 1556ndash1564 June 2015
[33] G Cheng C Yang X Yao L Guo and J Han ldquoWhenDeep Learning Meets Metric Learning Remote Sensing ImageScene Classification via Learning Discriminative CNNsrdquo IEEETransactions on Geoscience and Remote Sensing pp 1ndash11
[34] J He J Feng X Liu et al ldquoMobile product search with Bag ofHash Bits and boundary rerankingrdquo in Proceedings of the 2012IEEE Conference on Computer Vision and Pattern RecognitionCVPR 2012 pp 3005ndash3012 USA June 2012
Security and Communication Networks 13
[35] F Shen Y Mu Y Yang et al ldquoClassification by retrievalBinarizing data and classifiersrdquo in Proceedings of the 40thInternational ACM SIGIR Conference on Research and Develop-ment in Information Retrieval SIGIR 2017 pp 595ndash604 JapanAugust 2017
[36] P Li S Zhao andR Zhang ldquoA cluster analysis selection strategyfor supersaturated designsrdquo Computational Statistics amp DataAnalysis vol 54 no 6 pp 1605ndash1612 2010
[37] A Pradeep S Mridula and P Mohanan ldquoHigh securityidentity tags using spiral resonatorsrdquo Cmc-Computers Materialsamp Continua vol 52 no 3 pp 187ndash196 2016
[38] Y Cao Z Zhou X Sun and C Gao ldquoCoverless informationhiding based on the molecular structure images of materialrdquoComputers Materials and Continua vol 54 no 2 pp 197ndash2072018
[39] Y LiuH Peng and JWang ldquoVerifiable diversity ranking searchover encrypted outsourced datardquo Cmc-Computers Materials ampContinua vol 55 no 1 pp 037ndash057 2018
[40] K Lin J Lu C-S Chen and J Zhou ldquoLearning compactbinary descriptors with unsupervised deep neural networksrdquo inProceedings of the 2016 IEEEConference onComputer Vision andPattern Recognition CVPR 2016 pp 1183ndash1192 USA July 2016
[41] T Do A Doan and N Cheung ldquoLearning to Hash with BinaryDeep Neural Networkrdquo in Computer Vision ndash ECCV 2016vol 9909 of Lecture Notes in Computer Science pp 219ndash234Springer International Publishing Cham 2016
[42] Rui Zhang Di Xiao and Yanting Chang ldquoA Novel ImageAuthentication with Tamper Localization and Self-Recovery inEncrypted Domain Based on Compressive Sensingrdquo Securityand Communication Networks vol 2018 Article ID 1591206 15pages 2018
[43] Xia ShuangKui and JianbinWu ldquoAModification-Free Steganog-raphy Method Based on Image Information Entropyrdquo Securityand Communication Networks vol 2018 Article ID 6256872 8pages 2018
[44] J Zhang B Qu and N Xiu ldquoSome projection-like methods forthe generalized Nash equilibriardquo Computational Optimizationand Applications vol 45 no 1 pp 89ndash109 2010
[45] Biao Qu and Jing Zhao ldquoMethods for Solving Generalized NashEquilibriumrdquo Journal of Applied Mathematics vol 2013 ArticleID 762165 6 pages 2013
[46] CWang CMa and J Zhou ldquoA new class of exact penalty func-tions and penalty algorithmsrdquo Journal of Global Optimizationvol 58 no 1 pp 51ndash73 2014
[47] Y Wang X Sun and F Meng ldquoOn the conditional andpartial trade credit policywith capital constraints A StackelbergModelrdquo Applied Mathematical Modelling vol 40 no 1 pp 1ndash182016
[48] S Lian and Y Duan ldquoSmoothing of the lower-order exactpenalty function for inequality constrained optimizationrdquo Jour-nal of Inequalities and Applications Paper No 185 12 pages2016
Chemical EngineeringInternational Journal of Antennas and
Propagation
International Journal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Navigation and Observation
International Journal of
Hindawi
wwwhindawicom Volume 2018
Advances in
Multimedia
Submit your manuscripts atwwwhindawicom
6 Security and Communication Networks
StartTwo valued by the
Sigmod function in Hash layer
Building hash pool and hash bucket
Calculation of Hamming distance
Select the top m candidate images
Calculation of Euclidean distance
Select the top kimages from m
candidate imagesas retrieval result
End
Supervised pre-training
Rough image retrieval
If itgt05Yes
1
No
0
If it less than a threshold
Yes
No
Accurate image retrieval
The featureof the image
to be retrieved
Self-learning network
The featureof the
training images
Binary processing
Binary codes for trainingimage
Binary codes for retrieving
image
Figure 2 Deep learning-hash retrieval flowchart IDLHmainly includes three processes (supervised pretraining rough image retrieval andaccurate image retrieval)
+1+1 +1
+1
Layer 0Input layer
Layer 1 Layer 2 Layer 3 Hash Layer
x1
x2
x3
x4
x5
ℎ(1)1
ℎ(1)2
ℎ(1)3
ℎ(1)4
ℎ(2)1
ℎ(2)2
ℎ(2)3
ℎ(2)4
ℎ(3)1
ℎ(3)2
ℎ(3)3
Output ℎ(3)k
from layer 3 is usedas binary code with bk = f(ℎ(3)
k)
Figure 3 Self-learning network based on stack self-encoding network The neurons labeled xi is the input of the neural network and ldquo+ 1rdquoare the offset nodes (intercept entries) of the neural network The layer 0 is the input layer of neural network and layer 3 is the output layerof neural network The middle layers of layer 0 layer to layer 3 are the hidden layers of neural network
a small probability in events that is the nonadjacent data ishash to the same barrel Set the hash function as the follows
Here119908119896 is the projection vector and 119887119896 is the correspond-ing intercept The code value generated by formula (7) isminus1 1 and we use the following formula to convert it intotwo value codes
119910119896 = 12 (1 + ℎ119896 (119909)) (8)
Given a sample point isin 119877119863 we can compute a K-bitbinary code 119910 for 119909 with formula (9) The hash functionperforms the mapping as ℎ119896 119877119863 997888rarr 119861
Here 119883 = 119909119899119873119899=1 isin 119877119863times119873 is the feature data matrix withpoints as columns Such a binary encoding process can also beviewed as mapping the original data point to a binary valuedspace namely Hamming space
34 SimilarityMeasure After obtaining the binary hash codeof the image it is necessary tomeasure similarity between theretrieved image and the library image in the Hamming spaceThe smaller the Hamming distance is the closer distancebetween the two data is and the degree of similarity is higherotherwise the two data similarity is lower
Here oplus is an XOR operation The two sets of 119910119894 and 119910119895represent the hash code of the search image feature and theimage library is mapped through the hash function The newimage features learned by the stack self-encoding network aregenerated by the hash function The storage structure of theimage feature vectors is shown in Table 3
As can be seen from Table 3 the hash code of the imageis related to the image ID and the image name one by one Inthe process of searching the image feature vector is obtainedthrough deep learning by a hash function the original data ismapped into a newdata space and a corresponding hash codeis obtained The hash code is used to calculate the Hammingdistance in the Hamming space as a measure of similaritybetween images Finally the storage structure of the imagefeature vector is used to find the corresponding image ID ofthe hash code and the output retrieval result is output to theuser
35 Image Secondary Search In the first-level search phasethe features learned from the deep learning network aremapped into the Hamming space using the hash function Inthe similarity measurement phase the traditional Euclideandistance is abandoned Measure the similarity betweenimages by comparing the Hamming distance between theimage features of the query image and the image of thelibrary image In order to further improve the accuracy ofretrieval without affecting the real-time performance we canretrieve the image by the second level retrieval These stepsare described in detail as follows After one level retrievalwe choose the K images with the most similarity in thefirst-level retrieval result and then calculate the Euclideandistance between the original feature vector of the K imagesand the original feature vector of the query imageThe resultsobtained as the similarity measure of the images and outputthe retrieval result that has been ranked from the high andlow with the similarity distance
Although theHash algorithmmaps the high-dimensionalfeature vectors of the image into a hash-coded form theproblem of ldquodimensional disastersrdquo is solved and the retrievalefficiency is greatly accelerated However when the similaritycomparison is performed the Hamming distances of theimage features are simply compared using the results of theprimary search and occasionally undesirable results maystill appear on the search results If we want to increasethe accuracy of the search we must increase the hash codelength However excessively long codes will increase theamount of calculations increase the memory burden andreduce the real-time nature of retrieval failing to achieve
the goal of reducing the size of data In order to solve thisproblem keep the retrieval efficiency and further improvethe retrieval accuracy we propose a search strategy forsecondary retrieval the specific steps of which are as follows
Step 1 Through the first-level search in the Hamming spacethe similarity degree of the images is sorted and the top Ksorting images are selected
Step 2 For the 119896 images in Step 1 calculate the Euclideandistance one by one from its original image feature vector tothe image feature vector of the query image
Step 3 The Euclidean distance calculated in Step 2 is sortedThe smaller the calculated value is the higher similaritybetween images is and the similarity is sorted from high tolow and output as the final search result
In the second search it is necessary to pay attention tothe selection of the 119870 value although the larger the 119870 valueis the better the search effect is but accordingly the longerthe time is consumed Therefore it is necessary to combinevarious factors to select the appropriate119870 value
4 Experimental and Performance Analysis
In this section we thoroughly compare the proposedapproach with the improved deep learning hash retrievalmethods on several benchmark datasets Through a series ofexperiments the effectiveness and feasibility of the proposedalgorithm are verified
41 Database Two mostly used databases in the recent deeplearning hash works are taken into our evaluation Thetwo image libraries are derived from the CIFAR-10[11] coreexperimental image library dataset and theCaltech 256 imagelibrary dataset
CIFAR-10 dataset contains 10 object categories and eachclass consists of 6000 images resulting in a total of 60000images The dataset is split into training and test sets whichare averagely divided into 10 object classes The Caltech 256image library dataset contains 29780 color images which aregrouped intro 256 classes
First test the CIFAR-10 image dataset There are a totalof 50000 training samples which are used for training on thedeep learning networkThe remaining 10000 images are usedas test samples And then we randomly select 50 images fromdatabase as the query images For theHidden Image Retrievalalgorithm based on deep learning mentioned in this paperthe image pixel data is directly used as input while for otheralgorithms the 512-dimensional GIST feature is used as thefeature expression of the image Note quantization all imagesinto 32lowast32 sizes before experiment
For the Caltech 256 image a total of 256 classes areincluded and each class contains at least 70 images There-fore 70 images of each class a total of 17920 images arerandomly selected and are used as training images Theremaining images are used as test samples In addition all ofthe imagesrsquo size is set to 64lowast64 again when training
8 Security and Communication Networks
42 Evaluation Metrics We measure the performance ofcompared methods using Precision-recall and Average-Retrieval Precision (ARP) curves Precision is the ratio ofthe correct number of images m in the search result to thenumber k of all returned images The formula is as follows
precision = 119898119896 times 100 (12)
Recall is the ratio of the correct number of imagesm in thesearch results to the number g of images in the image libraryThe formula is as follows
recall = 119898119902 times 100 (13)
Assume that the search result of the query image 119894 is119861119894 and 119860 119894 means that the category is the same between thequery image and the return image then the accuracy rate forthe image query result 119875(119894) can be defined by the followingformula
119875 (119894) = |119860 (119894) cap 119861 (119894)||119861 (119894)| (14)
Average-Retrieval Precision (ARP) the average value ofall the images in the same class as the retrieval rate obtainedfrom the retrieval image is defined as follows
Here 119868119863119898 is the category index number of the image119898 isthe category index119873 is the number of imageswhose categoryis 119868119863119898 and 119894119889(119894) is the category index number of the queryimage
43 Performance Analysis In the proposed algorithm IDLHthe length of the hash sequence and the depth of the hiddenlayer in the deep learning network are two key parametersWhen the hash sequence length is small different featurevectors can easily be mapped into the same hash sequence sothe retrieval accuracy is low However if the hash sequence istoo long a large storage space is required and a long time isconsumed which reduces the real-time performance For thenumber of hidden layers the number of layers in the hiddenlayer is too small which is not conducive to learning strongimage features However if the depth of the hidden layer istoo large the difficulty of training is increased In order toverify the effectiveness and feasibility of our algorithm weconducted the following experiments
(1) Results on CIFAR-10 Dataset Figure 4(a) shows the searchresults of the Average-Retrieval Precision using our proposedalgorithm IDLH compared with the LSH algorithm [3] andother three deep learning algorithms the DH algorithm [21]the DeepBit algorithm [40] and the UH-BDNN algorithm[41] on CIFAR-10 dataset with 8 16 32 48 64 96 128and 256 bits Figure 4(b) shows the Precision-recall curveunder 48-bit encoding It can be seen that the algorithm hasa higher precision than the other hashing algorithms with
the same recall rate However the advantage is not obviousand the average accuracy is slightly higher than other hashalgorithms
In order to overcome the above defects we use deeplearning to perform hash mapping on image features per-form hash encoding of different bits on the same featurecalculate the Precision-recall of the search results under thecondition of different coded bits and determine the impactof the encoding length on the retrieval results
As Figure 5 shows with the increase in the number ofcoded bits the Precision-recall is continuously increasingWith the increase in the number of coded bits the imageis better expressed However after the number of coded bitsreaches 64 even if the number of coded bits increases theaverage accuracy rate increases relatively slowly Because theinformation of the tiny image is relatively simple when thenumber of encoded bits reaches 64 bits a relatively goodimage expression has been obtained and the performanceof the algorithm has basically stabilized At this time despiteincrease in the number of encoding bits it is not very helpfulto improve the accuracy rate
In addition we want to test the influence of the numberof hidden layers in the deep learning network on the retrievalresult by changing the number of hidden layers
Figure 6 shows the effect of deep learning networks onexperimental results in the case of different hidden layernumbers
It can be seen that deeper networks do not have muchimprovement in performance which is different from theexpectation that more hidden layers will help learn strongerimage features Since the image library data used in theexperiment is a tiny image library relatively good image char-acteristics can be learned using a deep learning network withfewer layers However if the image library is replaced witha more colorful image the deep neural network can acquiremore detailed image features and the deepening of thelearning network will greatly help the study of image features
(2) Results on Caltech 256 Image Data Set Figure 7(a)shows the results of the Average-Retrieval Precision resultswhen the number of coded bits is different Compared withthe black and white image library the proposed algorithmembodies the advantage of image feature learning and leadsthe Average-Retrieval Precision to other hash retrieval algo-rithms In Figure 7(b) we can also see that the algorithmproposed in this paper has a higher Precision-recall thanother algorithms under the same recall rate and it has bettersearch performance
As shown in Figure 8 as the number of coded bitsincreases the precision rate increases with the same recallrate This feature of the color image library is more pro-nounced than the black and white image library Becausethe color image contains more information more codingis needed to express it and the increase of encoding helpsto learn the features of the image The experimental resultsalso show that the deep learning network has learned moreexcellent image features
In Figure 9 the precision rate is significantly improvedby the increase in the number of hidden layers in the
Security and Communication Networks 9
IDLHLSHDHDeepBitUH-BDNN
16 32 48 64 96 128 2568Bits
0
01
02
03
04
05
06
07
08Av
erag
e Ret
rieva
l Pre
cisio
n
(a) Average-Retrieval Precision
IDLHLSHDHDeepBitUH-BDNN
01 02 03 04 05 06 07 08 09 100Recall
0
01
02
03
04
05
06
07
08
09
Prec
ision
(b) Precision-recall at 48 bits
Figure 4 Five kinds of algorithm retrieval performance comparison on CIFAR-10 dataset
8 bits16 bits32 bits48 bits64 bits128 bits
01 02 03 04 05 06 07 08 09 10Recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
Figure 5 Precision-recall curves with lengths
color image library Caltech 256 This is because theinformation contained in a more colorful image is morecomplex Adding a hidden layer can learn more detailsof the image and help improve the accuracy of thesearch
Next we tested the performance of secondary imageretrieval The value of k in the secondary search is 20 andthe experimental results are shown in Figure 10 As can
Level 1Level 2Level 3Level 4
01 02 03 04 05 06 07 08 09 1000Recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
Figure 6 Precision-recall curves with different code differenthidden layers
be seen from the results secondary retrieval can effectivelyimprove the retrieval accuracy when the number of codedbits is small However with the increase in the number ofencoding bits the results of the secondary search and theaccuracy of the primary search are not much different Thisis because the shorter the hash sequence is the easier thefeature vectors with different original features are mappedto the same hash code In order to make up for the errors
10 Security and Communication Networks
IDLHLSHDHDeepBitUH-BDNN
16 32 48 64 96 128 2568Bits
0
01
02
03
04
05
06
07
08
09
1Av
erag
e-Re
trie
val P
reci
sion
(a) Average-Retrieval Precision
IDLHLSHDHDeepBitUH-BDNN
01 02 03 04 05 06 07 08 09 1000Precision-recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
(b) Precision-recall at 48 bits
Figure 7 Five kinds of algorithm retrieval performance comparison on Caltech 256 set
8 bits16 bits32 bits48 bits64 bits128 bits
01 02 03 04 05 06 07 08 09 10Recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
Figure 8 Precision-recall curves with different code lengths
caused by the short hash code it is necessary to performsecondary searchWhen the number of encoding bits is smalla secondary retrieval method is used in IDLH and the searchaccuracy rate can be improved at the expense of a small searchspeed
Level 1Level 2Level 3Level 4
01 02 03 04 05 06 07 08 09 10Recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
Figure 9 Precision-recall curves with different hidden layers
5 Conclusion
With the rapid development of data storage and digital pro-cess more and more digital information is transformed andtransmitted over the Internet day by day which brings peoplea series of security problems as well as convenience [42] Theresearches on digital image security that is image encryptionimage data hiding and image authentication become moreimportant than ever [38 39] The most essential problem of
Security and Communication Networks 11
First retrievalSecond retrieval
16 32 48 64 96 128 2568Bits
07
08
09
1
Aver
age-
Retr
ieva
l Pre
cisio
n
Figure 10 Average-Retrieval Precision with first and secondretrieval
image recognition is to extract robust features The qualityof feature extraction is directly related to the effect of recog-nition so most of the previous work on image recognitionis spent on artificial design features [43] In recent yearsthe emergence of deep learning technology has changed thestatus of artificial design classification characteristics Deeplearning technology simulates the mechanism of humanvisual system information classification processing from themost primitive image pixels to lower edge features then tothe target components that are combined on the edge andfinally to the whole target depth learning can be combinedby layer by layer The high-level feature is the combination oflow level features From low level to high level features aremore and more abstract and show semantics more and moreFrom the underlying features to the combination of high-levelfeatures it is the depth of learning that is done by itself Itdoes not require manual intervention Compared with thecharacteristics of the artificial design this combination offeatures can be closer to the semantic expression
In terms of illegal image retrieval the traditional recog-nition method should establish a recognition model foreach type of recognition task In the actual application arecognition model needs a recognition server If there aremany identification tasks the cost is too high We used thedeep neural network to recognize the illegal image it onlyneeds to collect the samples of every kind of illegal imageand participate in the training of the deep neural networkFinally a multiclassification recognition model is trainedWhen classifying unknown samples deep neural networkaccounting calculates the probability that the image belongsto each class
We all know that in the image detection process theaccuracy and recall rate are mutually influential Ideally bothmust be high but in general the accuracy is high and the
recall rate is low the recall rate is high and the accuracy islow For image retrieval we need to improve the accuracyunder the condition of guaranteeing the recall rate Forimage disease surveillance and anti-illegal images we need toenhance the recall under the condition of ensuring accuracyTherefore in different application scenarios in order toachieve a balance between accuracy and recall perhaps somegame theory (such as Nash Equilibrium [44 45]) and penaltyfunction [46ndash48] can provide related optimization solutions
In this paper we proposed an improved deep-learning-hashing approach IDLH which optimized over two majorimage retrieval process
(a) In the feature extraction process the self-encodednetwork of the look-ahead type is trained by using unlabeledimage data and the expression of robust image features islearned This unlabeled learning method does not requireimage library labeling and reduces the requirements for theimage library At the same time it also takes advantage of thedeep learning networks strong learning ability and obtainsbetter image feature expression than ordinary algorithms
(b) On the index structure a secondary search is pro-posed which further increases the accuracy of the search atthe expense of very little retrieval time
Through experiments the algorithm proposed in thispaper is compared with other classic hashing algorithms onmultiple evaluation indicators Firstly we tested the learningnetworks of different code lengths and depths in order totest their effect on the retrieval system and then tested theperformance of the secondary search Through the above-mentioned series of experiments for different parameters theeffectiveness of the improved deep learning hash retrievalalgorithm proposed in this paper is verified and throughthe experimental data the good retrieval results are provedIn addition the proposed deep hashing training strategycan also be potentially applied to other hashing problemsinvolving data similarity computation
Data Availability
The data used to support the findings of this study areincluded within the article
Conflicts of Interest
The authors declare that they have no conflicts of interest
Acknowledgments
The work was funded by the National Natural ScienceFoundation of China (Grants nos 61206138 and 61373016)
References
[1] R Datta D Joshi J Li and J Z Wang ldquoImage retrieval ideasinfluences and trends of the new agerdquoACMComputing Surveysvol 40 no 2 article 5 2008
[2] G Shakhnarovich T Darrell and P Indyk Nearest-NeighborMethods in Learning andVisionTheory and PracticeMITPressCambridge MA USA 2006
12 Security and Communication Networks
[3] A Gionis P Indyk and R Motwani ldquoSimilarity search in highdimensions via hashingrdquo in 25th Int Conf pp 518ndash529 1999
[4] Z Pan J Lei Y Zhang and F L Wang ldquoAdaptive fractional-Pixel motion estimation skipped algorithm for efficient HEVCmotion estimationrdquoACMTransactions onMultimedia Comput-ing Communications and Applications (TOMM) vol 14 no 1pp 1ndash19 2018
[5] G-L Tian M Wang and L Song ldquoVariable selection in thehigh-dimensional continuous generalized linear model withcurrent status datardquo Journal of Applied Statistics vol 41 no 3pp 467ndash483 2014
[6] M Datar N Immorlica P Indyk and V S Mirrokni ldquoLocality-sensitive hashing scheme based on p-stable distributionsrdquo inProceedings of the 20th Annual Symposium on ComputationalGeometry (SCG rsquo04) pp 253ndash262 ACM June 2004
[7] B Kulis P Jain and K Grauman ldquoFast similarity search forlearned metricsrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 31 no 12 pp 2143ndash2157 2009
[8] M Raginsky and S Lazebnik ldquoLocality-sensitive binary codesfrom shift-invariant kernelsrdquo in Proceedings of the 23rd AnnualConference on Neural Information Processing Systems NIPS2009 pp 1509ndash1517 Canada December 2009
[9] L Qi X Zhang W Dou and Q Ni ldquoA distributed locality-sensitive hashing-based approach for cloud service recommen-dation from multi-source datardquo IEEE Journal on Selected Areasin Communications vol 35 no 11 pp 2616ndash2624 2017
[10] M A Carreira-Perpinan and R Raziperchikolaei ldquoHashingwith binary autoencodersrdquo in Proceedings of the IEEE Confer-ence on Computer Vision and Pattern Recognition CVPR 2015pp 557ndash566 USA June 2015
[11] J Wang S Kumar and S-F Chang ldquoSemi-supervised hashingfor large-scale searchrdquo IEEE Transactions on Pattern Analysisand Machine Intelligence vol 34 no 12 pp 2393ndash2406 2012
[12] Y Gong S Lazebnik and A Gordo ldquoIterative quantizationa Procrustean approach to learning binary codes for large-scale image retrievalrdquo in Proceedings of the IEEE Conference onComputer Vision and Pattern Recognition (CVPR rsquo11) pp 2916ndash2929 June 2011
[13] W Kong and W J Li ldquoIsotropic hashingrdquo NIPS vol 25 2012[14] M Norouzi and D J Fleet ldquoMinimal loss hashing for compact
binary codesrdquo in Proceedings of the 28th International Confer-ence on Machine Learning ICML 2011 pp 353ndash360 USA July2011
[15] J Wang W Liu A X Sun and Y-G Jiang ldquoLearning hashcodes with listwise supervisionrdquo in Proceedings of the 2013 14thIEEE International Conference on Computer Vision ICCV 2013pp 3032ndash3039 Australia December 2013
[16] G Lin C Shen Q Shi A Van Den Hengel and D Suter ldquoFastsupervised hashing with decision trees for high-dimensionaldatardquo in Proceedings of 27th IEEE Conference on ComputerVision and Pattern Recognition CVPRrsquo pp 1971ndash1978 USA2014
[17] Y Gong S Kumar H A Rowley and S Lazebnik ldquoLearningbinary codes for high-dimensional data using bilinear projec-tionsrdquo in Proceedings of the 26th IEEE Conference on ComputerVision and Pattern Recognition CVPR 2013 pp 484ndash491 USAJune 2013
[18] W Liu J Wang Y Mu and S Kumar ldquoCompact hyperplanehashing with bilinear functionsrdquo in The 29th InternationalConference on Machine Learning (ICML12) pp 467ndash474 2012
[19] Y Weiss A Torralba and R Fergus ldquoSpectral hashingrdquo inProceedings of the 22nd Annual Conference on Neural Informa-tion Processing Systems (NIPS rsquo08) pp 1753ndash1760 VancouverCanada December 2008
[20] W Liu J Wang S Kumar and S F Chang ldquoHashing withgraphsrdquo inThe 28th international conference on machine learn-ing (ICML11) 2011
[21] F Shen X Zhou Y Yang J Song H T Shen and D Tao ldquoA fastoptimization method for general binary code learningrdquo IEEETransactions on Image Processing vol 25 no 12 pp 5610ndash56212016
[22] F Shen W Liu S Zhang Y Yang and H T Shen ldquoLearningbinary codes for maximum inner product searchrdquo in Proceed-ings of the 15th IEEE International Conference on ComputerVision ICCV 2015 pp 4148ndash4156 Chile December 2015
[23] A Krizhevsky I Sutskever andG EHinton ldquoImagenet classifi-cation with deep convolutional neural networksrdquo in Proceedingsof the 26th Annual Conference on Neural Information ProcessingSystems (NIPS rsquo12) pp 1097ndash1105 Lake Tahoe Nev USADecember 2012
[24] G E Hinton and R R Salakhutdinov ldquoReducing the dimen-sionality of data with neural networksrdquo The American Associa-tion for the Advancement of Science Science vol 313 no 5786pp 504ndash507 2006
[25] A Torralba R Fergus and Y Weiss ldquoSmall codes and largeimage databases for recognitionrdquo in Proceedings of the IEEEComputer Society Conference on Computer Vision and PatternRecognition (CVPR rsquo08) pp 1ndash8 2008
[26] R Salakhutdinov andG Hinton ldquoLearning a nonlinear embed-ding by preserving class neighbourhood structurerdquo Journal ofMachine Learning Research vol 2 pp 412ndash419 2007
[27] V E Liong J Lu GWang P Moulin and J Zhou ldquoDeep hash-ing for compact binary codes learningrdquo in Proceedings of theIEEE Conference on Computer Vision and Pattern RecognitionCVPR 2015 pp 2475ndash2483 USA June 2015
[28] Y Gong S Lazebnik A Gordo and F Perronnin ldquoIterativequantization A procrustean approach to learning binary codesfor large-scale image retrievalrdquo IEEE Transactions on PatternAnalysis andMachine Intelligence vol 35 no 12 pp 2916ndash29292013
[29] W Liu J Wang R Ji Y-G Jiang and S-F Chang ldquoSupervisedhashing with kernelsrdquo in Proceedings of the IEEE Conference onComputer Vision and Pattern Recognition (CVPR rsquo12) pp 2074ndash2081 Providence RI USA June 2012
[30] J Masci A Bronstein M Bronstein and P SprechmannldquoSparse similarity-preserving hashingrdquo in Int Conf LearnRepresent pp 1ndash13 2014
[31] B Kulis and K Grauman ldquoKernelized locality-sensitive hash-ingrdquo IEEE Transactions on Pattern Analysis and Machine Intel-ligence vol 34 no 6 pp 1092ndash1104 2012
[32] F Zhao YHuang LWang and T Tan ldquoDeep semantic rankingbased hashing for multi-label image retrievalrdquo in Proceedings ofIEEE Conference on Computer Vision and Pattern RecognitionCVPR 2015 pp 1556ndash1564 June 2015
[33] G Cheng C Yang X Yao L Guo and J Han ldquoWhenDeep Learning Meets Metric Learning Remote Sensing ImageScene Classification via Learning Discriminative CNNsrdquo IEEETransactions on Geoscience and Remote Sensing pp 1ndash11
[34] J He J Feng X Liu et al ldquoMobile product search with Bag ofHash Bits and boundary rerankingrdquo in Proceedings of the 2012IEEE Conference on Computer Vision and Pattern RecognitionCVPR 2012 pp 3005ndash3012 USA June 2012
Security and Communication Networks 13
[35] F Shen Y Mu Y Yang et al ldquoClassification by retrievalBinarizing data and classifiersrdquo in Proceedings of the 40thInternational ACM SIGIR Conference on Research and Develop-ment in Information Retrieval SIGIR 2017 pp 595ndash604 JapanAugust 2017
[36] P Li S Zhao andR Zhang ldquoA cluster analysis selection strategyfor supersaturated designsrdquo Computational Statistics amp DataAnalysis vol 54 no 6 pp 1605ndash1612 2010
[37] A Pradeep S Mridula and P Mohanan ldquoHigh securityidentity tags using spiral resonatorsrdquo Cmc-Computers Materialsamp Continua vol 52 no 3 pp 187ndash196 2016
[38] Y Cao Z Zhou X Sun and C Gao ldquoCoverless informationhiding based on the molecular structure images of materialrdquoComputers Materials and Continua vol 54 no 2 pp 197ndash2072018
[39] Y LiuH Peng and JWang ldquoVerifiable diversity ranking searchover encrypted outsourced datardquo Cmc-Computers Materials ampContinua vol 55 no 1 pp 037ndash057 2018
[40] K Lin J Lu C-S Chen and J Zhou ldquoLearning compactbinary descriptors with unsupervised deep neural networksrdquo inProceedings of the 2016 IEEEConference onComputer Vision andPattern Recognition CVPR 2016 pp 1183ndash1192 USA July 2016
[41] T Do A Doan and N Cheung ldquoLearning to Hash with BinaryDeep Neural Networkrdquo in Computer Vision ndash ECCV 2016vol 9909 of Lecture Notes in Computer Science pp 219ndash234Springer International Publishing Cham 2016
[42] Rui Zhang Di Xiao and Yanting Chang ldquoA Novel ImageAuthentication with Tamper Localization and Self-Recovery inEncrypted Domain Based on Compressive Sensingrdquo Securityand Communication Networks vol 2018 Article ID 1591206 15pages 2018
[43] Xia ShuangKui and JianbinWu ldquoAModification-Free Steganog-raphy Method Based on Image Information Entropyrdquo Securityand Communication Networks vol 2018 Article ID 6256872 8pages 2018
[44] J Zhang B Qu and N Xiu ldquoSome projection-like methods forthe generalized Nash equilibriardquo Computational Optimizationand Applications vol 45 no 1 pp 89ndash109 2010
[45] Biao Qu and Jing Zhao ldquoMethods for Solving Generalized NashEquilibriumrdquo Journal of Applied Mathematics vol 2013 ArticleID 762165 6 pages 2013
[46] CWang CMa and J Zhou ldquoA new class of exact penalty func-tions and penalty algorithmsrdquo Journal of Global Optimizationvol 58 no 1 pp 51ndash73 2014
[47] Y Wang X Sun and F Meng ldquoOn the conditional andpartial trade credit policywith capital constraints A StackelbergModelrdquo Applied Mathematical Modelling vol 40 no 1 pp 1ndash182016
[48] S Lian and Y Duan ldquoSmoothing of the lower-order exactpenalty function for inequality constrained optimizationrdquo Jour-nal of Inequalities and Applications Paper No 185 12 pages2016
Here oplus is an XOR operation The two sets of 119910119894 and 119910119895represent the hash code of the search image feature and theimage library is mapped through the hash function The newimage features learned by the stack self-encoding network aregenerated by the hash function The storage structure of theimage feature vectors is shown in Table 3
As can be seen from Table 3 the hash code of the imageis related to the image ID and the image name one by one Inthe process of searching the image feature vector is obtainedthrough deep learning by a hash function the original data ismapped into a newdata space and a corresponding hash codeis obtained The hash code is used to calculate the Hammingdistance in the Hamming space as a measure of similaritybetween images Finally the storage structure of the imagefeature vector is used to find the corresponding image ID ofthe hash code and the output retrieval result is output to theuser
35 Image Secondary Search In the first-level search phasethe features learned from the deep learning network aremapped into the Hamming space using the hash function Inthe similarity measurement phase the traditional Euclideandistance is abandoned Measure the similarity betweenimages by comparing the Hamming distance between theimage features of the query image and the image of thelibrary image In order to further improve the accuracy ofretrieval without affecting the real-time performance we canretrieve the image by the second level retrieval These stepsare described in detail as follows After one level retrievalwe choose the K images with the most similarity in thefirst-level retrieval result and then calculate the Euclideandistance between the original feature vector of the K imagesand the original feature vector of the query imageThe resultsobtained as the similarity measure of the images and outputthe retrieval result that has been ranked from the high andlow with the similarity distance
Although theHash algorithmmaps the high-dimensionalfeature vectors of the image into a hash-coded form theproblem of ldquodimensional disastersrdquo is solved and the retrievalefficiency is greatly accelerated However when the similaritycomparison is performed the Hamming distances of theimage features are simply compared using the results of theprimary search and occasionally undesirable results maystill appear on the search results If we want to increasethe accuracy of the search we must increase the hash codelength However excessively long codes will increase theamount of calculations increase the memory burden andreduce the real-time nature of retrieval failing to achieve
the goal of reducing the size of data In order to solve thisproblem keep the retrieval efficiency and further improvethe retrieval accuracy we propose a search strategy forsecondary retrieval the specific steps of which are as follows
Step 1 Through the first-level search in the Hamming spacethe similarity degree of the images is sorted and the top Ksorting images are selected
Step 2 For the 119896 images in Step 1 calculate the Euclideandistance one by one from its original image feature vector tothe image feature vector of the query image
Step 3 The Euclidean distance calculated in Step 2 is sortedThe smaller the calculated value is the higher similaritybetween images is and the similarity is sorted from high tolow and output as the final search result
In the second search it is necessary to pay attention tothe selection of the 119870 value although the larger the 119870 valueis the better the search effect is but accordingly the longerthe time is consumed Therefore it is necessary to combinevarious factors to select the appropriate119870 value
4 Experimental and Performance Analysis
In this section we thoroughly compare the proposedapproach with the improved deep learning hash retrievalmethods on several benchmark datasets Through a series ofexperiments the effectiveness and feasibility of the proposedalgorithm are verified
41 Database Two mostly used databases in the recent deeplearning hash works are taken into our evaluation Thetwo image libraries are derived from the CIFAR-10[11] coreexperimental image library dataset and theCaltech 256 imagelibrary dataset
CIFAR-10 dataset contains 10 object categories and eachclass consists of 6000 images resulting in a total of 60000images The dataset is split into training and test sets whichare averagely divided into 10 object classes The Caltech 256image library dataset contains 29780 color images which aregrouped intro 256 classes
First test the CIFAR-10 image dataset There are a totalof 50000 training samples which are used for training on thedeep learning networkThe remaining 10000 images are usedas test samples And then we randomly select 50 images fromdatabase as the query images For theHidden Image Retrievalalgorithm based on deep learning mentioned in this paperthe image pixel data is directly used as input while for otheralgorithms the 512-dimensional GIST feature is used as thefeature expression of the image Note quantization all imagesinto 32lowast32 sizes before experiment
For the Caltech 256 image a total of 256 classes areincluded and each class contains at least 70 images There-fore 70 images of each class a total of 17920 images arerandomly selected and are used as training images Theremaining images are used as test samples In addition all ofthe imagesrsquo size is set to 64lowast64 again when training
8 Security and Communication Networks
42 Evaluation Metrics We measure the performance ofcompared methods using Precision-recall and Average-Retrieval Precision (ARP) curves Precision is the ratio ofthe correct number of images m in the search result to thenumber k of all returned images The formula is as follows
precision = 119898119896 times 100 (12)
Recall is the ratio of the correct number of imagesm in thesearch results to the number g of images in the image libraryThe formula is as follows
recall = 119898119902 times 100 (13)
Assume that the search result of the query image 119894 is119861119894 and 119860 119894 means that the category is the same between thequery image and the return image then the accuracy rate forthe image query result 119875(119894) can be defined by the followingformula
119875 (119894) = |119860 (119894) cap 119861 (119894)||119861 (119894)| (14)
Average-Retrieval Precision (ARP) the average value ofall the images in the same class as the retrieval rate obtainedfrom the retrieval image is defined as follows
Here 119868119863119898 is the category index number of the image119898 isthe category index119873 is the number of imageswhose categoryis 119868119863119898 and 119894119889(119894) is the category index number of the queryimage
43 Performance Analysis In the proposed algorithm IDLHthe length of the hash sequence and the depth of the hiddenlayer in the deep learning network are two key parametersWhen the hash sequence length is small different featurevectors can easily be mapped into the same hash sequence sothe retrieval accuracy is low However if the hash sequence istoo long a large storage space is required and a long time isconsumed which reduces the real-time performance For thenumber of hidden layers the number of layers in the hiddenlayer is too small which is not conducive to learning strongimage features However if the depth of the hidden layer istoo large the difficulty of training is increased In order toverify the effectiveness and feasibility of our algorithm weconducted the following experiments
(1) Results on CIFAR-10 Dataset Figure 4(a) shows the searchresults of the Average-Retrieval Precision using our proposedalgorithm IDLH compared with the LSH algorithm [3] andother three deep learning algorithms the DH algorithm [21]the DeepBit algorithm [40] and the UH-BDNN algorithm[41] on CIFAR-10 dataset with 8 16 32 48 64 96 128and 256 bits Figure 4(b) shows the Precision-recall curveunder 48-bit encoding It can be seen that the algorithm hasa higher precision than the other hashing algorithms with
the same recall rate However the advantage is not obviousand the average accuracy is slightly higher than other hashalgorithms
In order to overcome the above defects we use deeplearning to perform hash mapping on image features per-form hash encoding of different bits on the same featurecalculate the Precision-recall of the search results under thecondition of different coded bits and determine the impactof the encoding length on the retrieval results
As Figure 5 shows with the increase in the number ofcoded bits the Precision-recall is continuously increasingWith the increase in the number of coded bits the imageis better expressed However after the number of coded bitsreaches 64 even if the number of coded bits increases theaverage accuracy rate increases relatively slowly Because theinformation of the tiny image is relatively simple when thenumber of encoded bits reaches 64 bits a relatively goodimage expression has been obtained and the performanceof the algorithm has basically stabilized At this time despiteincrease in the number of encoding bits it is not very helpfulto improve the accuracy rate
In addition we want to test the influence of the numberof hidden layers in the deep learning network on the retrievalresult by changing the number of hidden layers
Figure 6 shows the effect of deep learning networks onexperimental results in the case of different hidden layernumbers
It can be seen that deeper networks do not have muchimprovement in performance which is different from theexpectation that more hidden layers will help learn strongerimage features Since the image library data used in theexperiment is a tiny image library relatively good image char-acteristics can be learned using a deep learning network withfewer layers However if the image library is replaced witha more colorful image the deep neural network can acquiremore detailed image features and the deepening of thelearning network will greatly help the study of image features
(2) Results on Caltech 256 Image Data Set Figure 7(a)shows the results of the Average-Retrieval Precision resultswhen the number of coded bits is different Compared withthe black and white image library the proposed algorithmembodies the advantage of image feature learning and leadsthe Average-Retrieval Precision to other hash retrieval algo-rithms In Figure 7(b) we can also see that the algorithmproposed in this paper has a higher Precision-recall thanother algorithms under the same recall rate and it has bettersearch performance
As shown in Figure 8 as the number of coded bitsincreases the precision rate increases with the same recallrate This feature of the color image library is more pro-nounced than the black and white image library Becausethe color image contains more information more codingis needed to express it and the increase of encoding helpsto learn the features of the image The experimental resultsalso show that the deep learning network has learned moreexcellent image features
In Figure 9 the precision rate is significantly improvedby the increase in the number of hidden layers in the
Security and Communication Networks 9
IDLHLSHDHDeepBitUH-BDNN
16 32 48 64 96 128 2568Bits
0
01
02
03
04
05
06
07
08Av
erag
e Ret
rieva
l Pre
cisio
n
(a) Average-Retrieval Precision
IDLHLSHDHDeepBitUH-BDNN
01 02 03 04 05 06 07 08 09 100Recall
0
01
02
03
04
05
06
07
08
09
Prec
ision
(b) Precision-recall at 48 bits
Figure 4 Five kinds of algorithm retrieval performance comparison on CIFAR-10 dataset
8 bits16 bits32 bits48 bits64 bits128 bits
01 02 03 04 05 06 07 08 09 10Recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
Figure 5 Precision-recall curves with lengths
color image library Caltech 256 This is because theinformation contained in a more colorful image is morecomplex Adding a hidden layer can learn more detailsof the image and help improve the accuracy of thesearch
Next we tested the performance of secondary imageretrieval The value of k in the secondary search is 20 andthe experimental results are shown in Figure 10 As can
Level 1Level 2Level 3Level 4
01 02 03 04 05 06 07 08 09 1000Recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
Figure 6 Precision-recall curves with different code differenthidden layers
be seen from the results secondary retrieval can effectivelyimprove the retrieval accuracy when the number of codedbits is small However with the increase in the number ofencoding bits the results of the secondary search and theaccuracy of the primary search are not much different Thisis because the shorter the hash sequence is the easier thefeature vectors with different original features are mappedto the same hash code In order to make up for the errors
10 Security and Communication Networks
IDLHLSHDHDeepBitUH-BDNN
16 32 48 64 96 128 2568Bits
0
01
02
03
04
05
06
07
08
09
1Av
erag
e-Re
trie
val P
reci
sion
(a) Average-Retrieval Precision
IDLHLSHDHDeepBitUH-BDNN
01 02 03 04 05 06 07 08 09 1000Precision-recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
(b) Precision-recall at 48 bits
Figure 7 Five kinds of algorithm retrieval performance comparison on Caltech 256 set
8 bits16 bits32 bits48 bits64 bits128 bits
01 02 03 04 05 06 07 08 09 10Recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
Figure 8 Precision-recall curves with different code lengths
caused by the short hash code it is necessary to performsecondary searchWhen the number of encoding bits is smalla secondary retrieval method is used in IDLH and the searchaccuracy rate can be improved at the expense of a small searchspeed
Level 1Level 2Level 3Level 4
01 02 03 04 05 06 07 08 09 10Recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
Figure 9 Precision-recall curves with different hidden layers
5 Conclusion
With the rapid development of data storage and digital pro-cess more and more digital information is transformed andtransmitted over the Internet day by day which brings peoplea series of security problems as well as convenience [42] Theresearches on digital image security that is image encryptionimage data hiding and image authentication become moreimportant than ever [38 39] The most essential problem of
Security and Communication Networks 11
First retrievalSecond retrieval
16 32 48 64 96 128 2568Bits
07
08
09
1
Aver
age-
Retr
ieva
l Pre
cisio
n
Figure 10 Average-Retrieval Precision with first and secondretrieval
image recognition is to extract robust features The qualityof feature extraction is directly related to the effect of recog-nition so most of the previous work on image recognitionis spent on artificial design features [43] In recent yearsthe emergence of deep learning technology has changed thestatus of artificial design classification characteristics Deeplearning technology simulates the mechanism of humanvisual system information classification processing from themost primitive image pixels to lower edge features then tothe target components that are combined on the edge andfinally to the whole target depth learning can be combinedby layer by layer The high-level feature is the combination oflow level features From low level to high level features aremore and more abstract and show semantics more and moreFrom the underlying features to the combination of high-levelfeatures it is the depth of learning that is done by itself Itdoes not require manual intervention Compared with thecharacteristics of the artificial design this combination offeatures can be closer to the semantic expression
In terms of illegal image retrieval the traditional recog-nition method should establish a recognition model foreach type of recognition task In the actual application arecognition model needs a recognition server If there aremany identification tasks the cost is too high We used thedeep neural network to recognize the illegal image it onlyneeds to collect the samples of every kind of illegal imageand participate in the training of the deep neural networkFinally a multiclassification recognition model is trainedWhen classifying unknown samples deep neural networkaccounting calculates the probability that the image belongsto each class
We all know that in the image detection process theaccuracy and recall rate are mutually influential Ideally bothmust be high but in general the accuracy is high and the
recall rate is low the recall rate is high and the accuracy islow For image retrieval we need to improve the accuracyunder the condition of guaranteeing the recall rate Forimage disease surveillance and anti-illegal images we need toenhance the recall under the condition of ensuring accuracyTherefore in different application scenarios in order toachieve a balance between accuracy and recall perhaps somegame theory (such as Nash Equilibrium [44 45]) and penaltyfunction [46ndash48] can provide related optimization solutions
In this paper we proposed an improved deep-learning-hashing approach IDLH which optimized over two majorimage retrieval process
(a) In the feature extraction process the self-encodednetwork of the look-ahead type is trained by using unlabeledimage data and the expression of robust image features islearned This unlabeled learning method does not requireimage library labeling and reduces the requirements for theimage library At the same time it also takes advantage of thedeep learning networks strong learning ability and obtainsbetter image feature expression than ordinary algorithms
(b) On the index structure a secondary search is pro-posed which further increases the accuracy of the search atthe expense of very little retrieval time
Through experiments the algorithm proposed in thispaper is compared with other classic hashing algorithms onmultiple evaluation indicators Firstly we tested the learningnetworks of different code lengths and depths in order totest their effect on the retrieval system and then tested theperformance of the secondary search Through the above-mentioned series of experiments for different parameters theeffectiveness of the improved deep learning hash retrievalalgorithm proposed in this paper is verified and throughthe experimental data the good retrieval results are provedIn addition the proposed deep hashing training strategycan also be potentially applied to other hashing problemsinvolving data similarity computation
Data Availability
The data used to support the findings of this study areincluded within the article
Conflicts of Interest
The authors declare that they have no conflicts of interest
Acknowledgments
The work was funded by the National Natural ScienceFoundation of China (Grants nos 61206138 and 61373016)
References
[1] R Datta D Joshi J Li and J Z Wang ldquoImage retrieval ideasinfluences and trends of the new agerdquoACMComputing Surveysvol 40 no 2 article 5 2008
[2] G Shakhnarovich T Darrell and P Indyk Nearest-NeighborMethods in Learning andVisionTheory and PracticeMITPressCambridge MA USA 2006
12 Security and Communication Networks
[3] A Gionis P Indyk and R Motwani ldquoSimilarity search in highdimensions via hashingrdquo in 25th Int Conf pp 518ndash529 1999
[4] Z Pan J Lei Y Zhang and F L Wang ldquoAdaptive fractional-Pixel motion estimation skipped algorithm for efficient HEVCmotion estimationrdquoACMTransactions onMultimedia Comput-ing Communications and Applications (TOMM) vol 14 no 1pp 1ndash19 2018
[5] G-L Tian M Wang and L Song ldquoVariable selection in thehigh-dimensional continuous generalized linear model withcurrent status datardquo Journal of Applied Statistics vol 41 no 3pp 467ndash483 2014
[6] M Datar N Immorlica P Indyk and V S Mirrokni ldquoLocality-sensitive hashing scheme based on p-stable distributionsrdquo inProceedings of the 20th Annual Symposium on ComputationalGeometry (SCG rsquo04) pp 253ndash262 ACM June 2004
[7] B Kulis P Jain and K Grauman ldquoFast similarity search forlearned metricsrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 31 no 12 pp 2143ndash2157 2009
[8] M Raginsky and S Lazebnik ldquoLocality-sensitive binary codesfrom shift-invariant kernelsrdquo in Proceedings of the 23rd AnnualConference on Neural Information Processing Systems NIPS2009 pp 1509ndash1517 Canada December 2009
[9] L Qi X Zhang W Dou and Q Ni ldquoA distributed locality-sensitive hashing-based approach for cloud service recommen-dation from multi-source datardquo IEEE Journal on Selected Areasin Communications vol 35 no 11 pp 2616ndash2624 2017
[10] M A Carreira-Perpinan and R Raziperchikolaei ldquoHashingwith binary autoencodersrdquo in Proceedings of the IEEE Confer-ence on Computer Vision and Pattern Recognition CVPR 2015pp 557ndash566 USA June 2015
[11] J Wang S Kumar and S-F Chang ldquoSemi-supervised hashingfor large-scale searchrdquo IEEE Transactions on Pattern Analysisand Machine Intelligence vol 34 no 12 pp 2393ndash2406 2012
[12] Y Gong S Lazebnik and A Gordo ldquoIterative quantizationa Procrustean approach to learning binary codes for large-scale image retrievalrdquo in Proceedings of the IEEE Conference onComputer Vision and Pattern Recognition (CVPR rsquo11) pp 2916ndash2929 June 2011
[13] W Kong and W J Li ldquoIsotropic hashingrdquo NIPS vol 25 2012[14] M Norouzi and D J Fleet ldquoMinimal loss hashing for compact
binary codesrdquo in Proceedings of the 28th International Confer-ence on Machine Learning ICML 2011 pp 353ndash360 USA July2011
[15] J Wang W Liu A X Sun and Y-G Jiang ldquoLearning hashcodes with listwise supervisionrdquo in Proceedings of the 2013 14thIEEE International Conference on Computer Vision ICCV 2013pp 3032ndash3039 Australia December 2013
[16] G Lin C Shen Q Shi A Van Den Hengel and D Suter ldquoFastsupervised hashing with decision trees for high-dimensionaldatardquo in Proceedings of 27th IEEE Conference on ComputerVision and Pattern Recognition CVPRrsquo pp 1971ndash1978 USA2014
[17] Y Gong S Kumar H A Rowley and S Lazebnik ldquoLearningbinary codes for high-dimensional data using bilinear projec-tionsrdquo in Proceedings of the 26th IEEE Conference on ComputerVision and Pattern Recognition CVPR 2013 pp 484ndash491 USAJune 2013
[18] W Liu J Wang Y Mu and S Kumar ldquoCompact hyperplanehashing with bilinear functionsrdquo in The 29th InternationalConference on Machine Learning (ICML12) pp 467ndash474 2012
[19] Y Weiss A Torralba and R Fergus ldquoSpectral hashingrdquo inProceedings of the 22nd Annual Conference on Neural Informa-tion Processing Systems (NIPS rsquo08) pp 1753ndash1760 VancouverCanada December 2008
[20] W Liu J Wang S Kumar and S F Chang ldquoHashing withgraphsrdquo inThe 28th international conference on machine learn-ing (ICML11) 2011
[21] F Shen X Zhou Y Yang J Song H T Shen and D Tao ldquoA fastoptimization method for general binary code learningrdquo IEEETransactions on Image Processing vol 25 no 12 pp 5610ndash56212016
[22] F Shen W Liu S Zhang Y Yang and H T Shen ldquoLearningbinary codes for maximum inner product searchrdquo in Proceed-ings of the 15th IEEE International Conference on ComputerVision ICCV 2015 pp 4148ndash4156 Chile December 2015
[23] A Krizhevsky I Sutskever andG EHinton ldquoImagenet classifi-cation with deep convolutional neural networksrdquo in Proceedingsof the 26th Annual Conference on Neural Information ProcessingSystems (NIPS rsquo12) pp 1097ndash1105 Lake Tahoe Nev USADecember 2012
[24] G E Hinton and R R Salakhutdinov ldquoReducing the dimen-sionality of data with neural networksrdquo The American Associa-tion for the Advancement of Science Science vol 313 no 5786pp 504ndash507 2006
[25] A Torralba R Fergus and Y Weiss ldquoSmall codes and largeimage databases for recognitionrdquo in Proceedings of the IEEEComputer Society Conference on Computer Vision and PatternRecognition (CVPR rsquo08) pp 1ndash8 2008
[26] R Salakhutdinov andG Hinton ldquoLearning a nonlinear embed-ding by preserving class neighbourhood structurerdquo Journal ofMachine Learning Research vol 2 pp 412ndash419 2007
[27] V E Liong J Lu GWang P Moulin and J Zhou ldquoDeep hash-ing for compact binary codes learningrdquo in Proceedings of theIEEE Conference on Computer Vision and Pattern RecognitionCVPR 2015 pp 2475ndash2483 USA June 2015
[28] Y Gong S Lazebnik A Gordo and F Perronnin ldquoIterativequantization A procrustean approach to learning binary codesfor large-scale image retrievalrdquo IEEE Transactions on PatternAnalysis andMachine Intelligence vol 35 no 12 pp 2916ndash29292013
[29] W Liu J Wang R Ji Y-G Jiang and S-F Chang ldquoSupervisedhashing with kernelsrdquo in Proceedings of the IEEE Conference onComputer Vision and Pattern Recognition (CVPR rsquo12) pp 2074ndash2081 Providence RI USA June 2012
[30] J Masci A Bronstein M Bronstein and P SprechmannldquoSparse similarity-preserving hashingrdquo in Int Conf LearnRepresent pp 1ndash13 2014
[31] B Kulis and K Grauman ldquoKernelized locality-sensitive hash-ingrdquo IEEE Transactions on Pattern Analysis and Machine Intel-ligence vol 34 no 6 pp 1092ndash1104 2012
[32] F Zhao YHuang LWang and T Tan ldquoDeep semantic rankingbased hashing for multi-label image retrievalrdquo in Proceedings ofIEEE Conference on Computer Vision and Pattern RecognitionCVPR 2015 pp 1556ndash1564 June 2015
[33] G Cheng C Yang X Yao L Guo and J Han ldquoWhenDeep Learning Meets Metric Learning Remote Sensing ImageScene Classification via Learning Discriminative CNNsrdquo IEEETransactions on Geoscience and Remote Sensing pp 1ndash11
[34] J He J Feng X Liu et al ldquoMobile product search with Bag ofHash Bits and boundary rerankingrdquo in Proceedings of the 2012IEEE Conference on Computer Vision and Pattern RecognitionCVPR 2012 pp 3005ndash3012 USA June 2012
Security and Communication Networks 13
[35] F Shen Y Mu Y Yang et al ldquoClassification by retrievalBinarizing data and classifiersrdquo in Proceedings of the 40thInternational ACM SIGIR Conference on Research and Develop-ment in Information Retrieval SIGIR 2017 pp 595ndash604 JapanAugust 2017
[36] P Li S Zhao andR Zhang ldquoA cluster analysis selection strategyfor supersaturated designsrdquo Computational Statistics amp DataAnalysis vol 54 no 6 pp 1605ndash1612 2010
[37] A Pradeep S Mridula and P Mohanan ldquoHigh securityidentity tags using spiral resonatorsrdquo Cmc-Computers Materialsamp Continua vol 52 no 3 pp 187ndash196 2016
[38] Y Cao Z Zhou X Sun and C Gao ldquoCoverless informationhiding based on the molecular structure images of materialrdquoComputers Materials and Continua vol 54 no 2 pp 197ndash2072018
[39] Y LiuH Peng and JWang ldquoVerifiable diversity ranking searchover encrypted outsourced datardquo Cmc-Computers Materials ampContinua vol 55 no 1 pp 037ndash057 2018
[40] K Lin J Lu C-S Chen and J Zhou ldquoLearning compactbinary descriptors with unsupervised deep neural networksrdquo inProceedings of the 2016 IEEEConference onComputer Vision andPattern Recognition CVPR 2016 pp 1183ndash1192 USA July 2016
[41] T Do A Doan and N Cheung ldquoLearning to Hash with BinaryDeep Neural Networkrdquo in Computer Vision ndash ECCV 2016vol 9909 of Lecture Notes in Computer Science pp 219ndash234Springer International Publishing Cham 2016
[42] Rui Zhang Di Xiao and Yanting Chang ldquoA Novel ImageAuthentication with Tamper Localization and Self-Recovery inEncrypted Domain Based on Compressive Sensingrdquo Securityand Communication Networks vol 2018 Article ID 1591206 15pages 2018
[43] Xia ShuangKui and JianbinWu ldquoAModification-Free Steganog-raphy Method Based on Image Information Entropyrdquo Securityand Communication Networks vol 2018 Article ID 6256872 8pages 2018
[44] J Zhang B Qu and N Xiu ldquoSome projection-like methods forthe generalized Nash equilibriardquo Computational Optimizationand Applications vol 45 no 1 pp 89ndash109 2010
[45] Biao Qu and Jing Zhao ldquoMethods for Solving Generalized NashEquilibriumrdquo Journal of Applied Mathematics vol 2013 ArticleID 762165 6 pages 2013
[46] CWang CMa and J Zhou ldquoA new class of exact penalty func-tions and penalty algorithmsrdquo Journal of Global Optimizationvol 58 no 1 pp 51ndash73 2014
[47] Y Wang X Sun and F Meng ldquoOn the conditional andpartial trade credit policywith capital constraints A StackelbergModelrdquo Applied Mathematical Modelling vol 40 no 1 pp 1ndash182016
[48] S Lian and Y Duan ldquoSmoothing of the lower-order exactpenalty function for inequality constrained optimizationrdquo Jour-nal of Inequalities and Applications Paper No 185 12 pages2016
Chemical EngineeringInternational Journal of Antennas and
Propagation
International Journal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Navigation and Observation
International Journal of
Hindawi
wwwhindawicom Volume 2018
Advances in
Multimedia
Submit your manuscripts atwwwhindawicom
8 Security and Communication Networks
42 Evaluation Metrics We measure the performance ofcompared methods using Precision-recall and Average-Retrieval Precision (ARP) curves Precision is the ratio ofthe correct number of images m in the search result to thenumber k of all returned images The formula is as follows
precision = 119898119896 times 100 (12)
Recall is the ratio of the correct number of imagesm in thesearch results to the number g of images in the image libraryThe formula is as follows
recall = 119898119902 times 100 (13)
Assume that the search result of the query image 119894 is119861119894 and 119860 119894 means that the category is the same between thequery image and the return image then the accuracy rate forthe image query result 119875(119894) can be defined by the followingformula
119875 (119894) = |119860 (119894) cap 119861 (119894)||119861 (119894)| (14)
Average-Retrieval Precision (ARP) the average value ofall the images in the same class as the retrieval rate obtainedfrom the retrieval image is defined as follows
Here 119868119863119898 is the category index number of the image119898 isthe category index119873 is the number of imageswhose categoryis 119868119863119898 and 119894119889(119894) is the category index number of the queryimage
43 Performance Analysis In the proposed algorithm IDLHthe length of the hash sequence and the depth of the hiddenlayer in the deep learning network are two key parametersWhen the hash sequence length is small different featurevectors can easily be mapped into the same hash sequence sothe retrieval accuracy is low However if the hash sequence istoo long a large storage space is required and a long time isconsumed which reduces the real-time performance For thenumber of hidden layers the number of layers in the hiddenlayer is too small which is not conducive to learning strongimage features However if the depth of the hidden layer istoo large the difficulty of training is increased In order toverify the effectiveness and feasibility of our algorithm weconducted the following experiments
(1) Results on CIFAR-10 Dataset Figure 4(a) shows the searchresults of the Average-Retrieval Precision using our proposedalgorithm IDLH compared with the LSH algorithm [3] andother three deep learning algorithms the DH algorithm [21]the DeepBit algorithm [40] and the UH-BDNN algorithm[41] on CIFAR-10 dataset with 8 16 32 48 64 96 128and 256 bits Figure 4(b) shows the Precision-recall curveunder 48-bit encoding It can be seen that the algorithm hasa higher precision than the other hashing algorithms with
the same recall rate However the advantage is not obviousand the average accuracy is slightly higher than other hashalgorithms
In order to overcome the above defects we use deeplearning to perform hash mapping on image features per-form hash encoding of different bits on the same featurecalculate the Precision-recall of the search results under thecondition of different coded bits and determine the impactof the encoding length on the retrieval results
As Figure 5 shows with the increase in the number ofcoded bits the Precision-recall is continuously increasingWith the increase in the number of coded bits the imageis better expressed However after the number of coded bitsreaches 64 even if the number of coded bits increases theaverage accuracy rate increases relatively slowly Because theinformation of the tiny image is relatively simple when thenumber of encoded bits reaches 64 bits a relatively goodimage expression has been obtained and the performanceof the algorithm has basically stabilized At this time despiteincrease in the number of encoding bits it is not very helpfulto improve the accuracy rate
In addition we want to test the influence of the numberof hidden layers in the deep learning network on the retrievalresult by changing the number of hidden layers
Figure 6 shows the effect of deep learning networks onexperimental results in the case of different hidden layernumbers
It can be seen that deeper networks do not have muchimprovement in performance which is different from theexpectation that more hidden layers will help learn strongerimage features Since the image library data used in theexperiment is a tiny image library relatively good image char-acteristics can be learned using a deep learning network withfewer layers However if the image library is replaced witha more colorful image the deep neural network can acquiremore detailed image features and the deepening of thelearning network will greatly help the study of image features
(2) Results on Caltech 256 Image Data Set Figure 7(a)shows the results of the Average-Retrieval Precision resultswhen the number of coded bits is different Compared withthe black and white image library the proposed algorithmembodies the advantage of image feature learning and leadsthe Average-Retrieval Precision to other hash retrieval algo-rithms In Figure 7(b) we can also see that the algorithmproposed in this paper has a higher Precision-recall thanother algorithms under the same recall rate and it has bettersearch performance
As shown in Figure 8 as the number of coded bitsincreases the precision rate increases with the same recallrate This feature of the color image library is more pro-nounced than the black and white image library Becausethe color image contains more information more codingis needed to express it and the increase of encoding helpsto learn the features of the image The experimental resultsalso show that the deep learning network has learned moreexcellent image features
In Figure 9 the precision rate is significantly improvedby the increase in the number of hidden layers in the
Security and Communication Networks 9
IDLHLSHDHDeepBitUH-BDNN
16 32 48 64 96 128 2568Bits
0
01
02
03
04
05
06
07
08Av
erag
e Ret
rieva
l Pre
cisio
n
(a) Average-Retrieval Precision
IDLHLSHDHDeepBitUH-BDNN
01 02 03 04 05 06 07 08 09 100Recall
0
01
02
03
04
05
06
07
08
09
Prec
ision
(b) Precision-recall at 48 bits
Figure 4 Five kinds of algorithm retrieval performance comparison on CIFAR-10 dataset
8 bits16 bits32 bits48 bits64 bits128 bits
01 02 03 04 05 06 07 08 09 10Recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
Figure 5 Precision-recall curves with lengths
color image library Caltech 256 This is because theinformation contained in a more colorful image is morecomplex Adding a hidden layer can learn more detailsof the image and help improve the accuracy of thesearch
Next we tested the performance of secondary imageretrieval The value of k in the secondary search is 20 andthe experimental results are shown in Figure 10 As can
Level 1Level 2Level 3Level 4
01 02 03 04 05 06 07 08 09 1000Recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
Figure 6 Precision-recall curves with different code differenthidden layers
be seen from the results secondary retrieval can effectivelyimprove the retrieval accuracy when the number of codedbits is small However with the increase in the number ofencoding bits the results of the secondary search and theaccuracy of the primary search are not much different Thisis because the shorter the hash sequence is the easier thefeature vectors with different original features are mappedto the same hash code In order to make up for the errors
10 Security and Communication Networks
IDLHLSHDHDeepBitUH-BDNN
16 32 48 64 96 128 2568Bits
0
01
02
03
04
05
06
07
08
09
1Av
erag
e-Re
trie
val P
reci
sion
(a) Average-Retrieval Precision
IDLHLSHDHDeepBitUH-BDNN
01 02 03 04 05 06 07 08 09 1000Precision-recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
(b) Precision-recall at 48 bits
Figure 7 Five kinds of algorithm retrieval performance comparison on Caltech 256 set
8 bits16 bits32 bits48 bits64 bits128 bits
01 02 03 04 05 06 07 08 09 10Recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
Figure 8 Precision-recall curves with different code lengths
caused by the short hash code it is necessary to performsecondary searchWhen the number of encoding bits is smalla secondary retrieval method is used in IDLH and the searchaccuracy rate can be improved at the expense of a small searchspeed
Level 1Level 2Level 3Level 4
01 02 03 04 05 06 07 08 09 10Recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
Figure 9 Precision-recall curves with different hidden layers
5 Conclusion
With the rapid development of data storage and digital pro-cess more and more digital information is transformed andtransmitted over the Internet day by day which brings peoplea series of security problems as well as convenience [42] Theresearches on digital image security that is image encryptionimage data hiding and image authentication become moreimportant than ever [38 39] The most essential problem of
Security and Communication Networks 11
First retrievalSecond retrieval
16 32 48 64 96 128 2568Bits
07
08
09
1
Aver
age-
Retr
ieva
l Pre
cisio
n
Figure 10 Average-Retrieval Precision with first and secondretrieval
image recognition is to extract robust features The qualityof feature extraction is directly related to the effect of recog-nition so most of the previous work on image recognitionis spent on artificial design features [43] In recent yearsthe emergence of deep learning technology has changed thestatus of artificial design classification characteristics Deeplearning technology simulates the mechanism of humanvisual system information classification processing from themost primitive image pixels to lower edge features then tothe target components that are combined on the edge andfinally to the whole target depth learning can be combinedby layer by layer The high-level feature is the combination oflow level features From low level to high level features aremore and more abstract and show semantics more and moreFrom the underlying features to the combination of high-levelfeatures it is the depth of learning that is done by itself Itdoes not require manual intervention Compared with thecharacteristics of the artificial design this combination offeatures can be closer to the semantic expression
In terms of illegal image retrieval the traditional recog-nition method should establish a recognition model foreach type of recognition task In the actual application arecognition model needs a recognition server If there aremany identification tasks the cost is too high We used thedeep neural network to recognize the illegal image it onlyneeds to collect the samples of every kind of illegal imageand participate in the training of the deep neural networkFinally a multiclassification recognition model is trainedWhen classifying unknown samples deep neural networkaccounting calculates the probability that the image belongsto each class
We all know that in the image detection process theaccuracy and recall rate are mutually influential Ideally bothmust be high but in general the accuracy is high and the
recall rate is low the recall rate is high and the accuracy islow For image retrieval we need to improve the accuracyunder the condition of guaranteeing the recall rate Forimage disease surveillance and anti-illegal images we need toenhance the recall under the condition of ensuring accuracyTherefore in different application scenarios in order toachieve a balance between accuracy and recall perhaps somegame theory (such as Nash Equilibrium [44 45]) and penaltyfunction [46ndash48] can provide related optimization solutions
In this paper we proposed an improved deep-learning-hashing approach IDLH which optimized over two majorimage retrieval process
(a) In the feature extraction process the self-encodednetwork of the look-ahead type is trained by using unlabeledimage data and the expression of robust image features islearned This unlabeled learning method does not requireimage library labeling and reduces the requirements for theimage library At the same time it also takes advantage of thedeep learning networks strong learning ability and obtainsbetter image feature expression than ordinary algorithms
(b) On the index structure a secondary search is pro-posed which further increases the accuracy of the search atthe expense of very little retrieval time
Through experiments the algorithm proposed in thispaper is compared with other classic hashing algorithms onmultiple evaluation indicators Firstly we tested the learningnetworks of different code lengths and depths in order totest their effect on the retrieval system and then tested theperformance of the secondary search Through the above-mentioned series of experiments for different parameters theeffectiveness of the improved deep learning hash retrievalalgorithm proposed in this paper is verified and throughthe experimental data the good retrieval results are provedIn addition the proposed deep hashing training strategycan also be potentially applied to other hashing problemsinvolving data similarity computation
Data Availability
The data used to support the findings of this study areincluded within the article
Conflicts of Interest
The authors declare that they have no conflicts of interest
Acknowledgments
The work was funded by the National Natural ScienceFoundation of China (Grants nos 61206138 and 61373016)
References
[1] R Datta D Joshi J Li and J Z Wang ldquoImage retrieval ideasinfluences and trends of the new agerdquoACMComputing Surveysvol 40 no 2 article 5 2008
[2] G Shakhnarovich T Darrell and P Indyk Nearest-NeighborMethods in Learning andVisionTheory and PracticeMITPressCambridge MA USA 2006
12 Security and Communication Networks
[3] A Gionis P Indyk and R Motwani ldquoSimilarity search in highdimensions via hashingrdquo in 25th Int Conf pp 518ndash529 1999
[4] Z Pan J Lei Y Zhang and F L Wang ldquoAdaptive fractional-Pixel motion estimation skipped algorithm for efficient HEVCmotion estimationrdquoACMTransactions onMultimedia Comput-ing Communications and Applications (TOMM) vol 14 no 1pp 1ndash19 2018
[5] G-L Tian M Wang and L Song ldquoVariable selection in thehigh-dimensional continuous generalized linear model withcurrent status datardquo Journal of Applied Statistics vol 41 no 3pp 467ndash483 2014
[6] M Datar N Immorlica P Indyk and V S Mirrokni ldquoLocality-sensitive hashing scheme based on p-stable distributionsrdquo inProceedings of the 20th Annual Symposium on ComputationalGeometry (SCG rsquo04) pp 253ndash262 ACM June 2004
[7] B Kulis P Jain and K Grauman ldquoFast similarity search forlearned metricsrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 31 no 12 pp 2143ndash2157 2009
[8] M Raginsky and S Lazebnik ldquoLocality-sensitive binary codesfrom shift-invariant kernelsrdquo in Proceedings of the 23rd AnnualConference on Neural Information Processing Systems NIPS2009 pp 1509ndash1517 Canada December 2009
[9] L Qi X Zhang W Dou and Q Ni ldquoA distributed locality-sensitive hashing-based approach for cloud service recommen-dation from multi-source datardquo IEEE Journal on Selected Areasin Communications vol 35 no 11 pp 2616ndash2624 2017
[10] M A Carreira-Perpinan and R Raziperchikolaei ldquoHashingwith binary autoencodersrdquo in Proceedings of the IEEE Confer-ence on Computer Vision and Pattern Recognition CVPR 2015pp 557ndash566 USA June 2015
[11] J Wang S Kumar and S-F Chang ldquoSemi-supervised hashingfor large-scale searchrdquo IEEE Transactions on Pattern Analysisand Machine Intelligence vol 34 no 12 pp 2393ndash2406 2012
[12] Y Gong S Lazebnik and A Gordo ldquoIterative quantizationa Procrustean approach to learning binary codes for large-scale image retrievalrdquo in Proceedings of the IEEE Conference onComputer Vision and Pattern Recognition (CVPR rsquo11) pp 2916ndash2929 June 2011
[13] W Kong and W J Li ldquoIsotropic hashingrdquo NIPS vol 25 2012[14] M Norouzi and D J Fleet ldquoMinimal loss hashing for compact
binary codesrdquo in Proceedings of the 28th International Confer-ence on Machine Learning ICML 2011 pp 353ndash360 USA July2011
[15] J Wang W Liu A X Sun and Y-G Jiang ldquoLearning hashcodes with listwise supervisionrdquo in Proceedings of the 2013 14thIEEE International Conference on Computer Vision ICCV 2013pp 3032ndash3039 Australia December 2013
[16] G Lin C Shen Q Shi A Van Den Hengel and D Suter ldquoFastsupervised hashing with decision trees for high-dimensionaldatardquo in Proceedings of 27th IEEE Conference on ComputerVision and Pattern Recognition CVPRrsquo pp 1971ndash1978 USA2014
[17] Y Gong S Kumar H A Rowley and S Lazebnik ldquoLearningbinary codes for high-dimensional data using bilinear projec-tionsrdquo in Proceedings of the 26th IEEE Conference on ComputerVision and Pattern Recognition CVPR 2013 pp 484ndash491 USAJune 2013
[18] W Liu J Wang Y Mu and S Kumar ldquoCompact hyperplanehashing with bilinear functionsrdquo in The 29th InternationalConference on Machine Learning (ICML12) pp 467ndash474 2012
[19] Y Weiss A Torralba and R Fergus ldquoSpectral hashingrdquo inProceedings of the 22nd Annual Conference on Neural Informa-tion Processing Systems (NIPS rsquo08) pp 1753ndash1760 VancouverCanada December 2008
[20] W Liu J Wang S Kumar and S F Chang ldquoHashing withgraphsrdquo inThe 28th international conference on machine learn-ing (ICML11) 2011
[21] F Shen X Zhou Y Yang J Song H T Shen and D Tao ldquoA fastoptimization method for general binary code learningrdquo IEEETransactions on Image Processing vol 25 no 12 pp 5610ndash56212016
[22] F Shen W Liu S Zhang Y Yang and H T Shen ldquoLearningbinary codes for maximum inner product searchrdquo in Proceed-ings of the 15th IEEE International Conference on ComputerVision ICCV 2015 pp 4148ndash4156 Chile December 2015
[23] A Krizhevsky I Sutskever andG EHinton ldquoImagenet classifi-cation with deep convolutional neural networksrdquo in Proceedingsof the 26th Annual Conference on Neural Information ProcessingSystems (NIPS rsquo12) pp 1097ndash1105 Lake Tahoe Nev USADecember 2012
[24] G E Hinton and R R Salakhutdinov ldquoReducing the dimen-sionality of data with neural networksrdquo The American Associa-tion for the Advancement of Science Science vol 313 no 5786pp 504ndash507 2006
[25] A Torralba R Fergus and Y Weiss ldquoSmall codes and largeimage databases for recognitionrdquo in Proceedings of the IEEEComputer Society Conference on Computer Vision and PatternRecognition (CVPR rsquo08) pp 1ndash8 2008
[26] R Salakhutdinov andG Hinton ldquoLearning a nonlinear embed-ding by preserving class neighbourhood structurerdquo Journal ofMachine Learning Research vol 2 pp 412ndash419 2007
[27] V E Liong J Lu GWang P Moulin and J Zhou ldquoDeep hash-ing for compact binary codes learningrdquo in Proceedings of theIEEE Conference on Computer Vision and Pattern RecognitionCVPR 2015 pp 2475ndash2483 USA June 2015
[28] Y Gong S Lazebnik A Gordo and F Perronnin ldquoIterativequantization A procrustean approach to learning binary codesfor large-scale image retrievalrdquo IEEE Transactions on PatternAnalysis andMachine Intelligence vol 35 no 12 pp 2916ndash29292013
[29] W Liu J Wang R Ji Y-G Jiang and S-F Chang ldquoSupervisedhashing with kernelsrdquo in Proceedings of the IEEE Conference onComputer Vision and Pattern Recognition (CVPR rsquo12) pp 2074ndash2081 Providence RI USA June 2012
[30] J Masci A Bronstein M Bronstein and P SprechmannldquoSparse similarity-preserving hashingrdquo in Int Conf LearnRepresent pp 1ndash13 2014
[31] B Kulis and K Grauman ldquoKernelized locality-sensitive hash-ingrdquo IEEE Transactions on Pattern Analysis and Machine Intel-ligence vol 34 no 6 pp 1092ndash1104 2012
[32] F Zhao YHuang LWang and T Tan ldquoDeep semantic rankingbased hashing for multi-label image retrievalrdquo in Proceedings ofIEEE Conference on Computer Vision and Pattern RecognitionCVPR 2015 pp 1556ndash1564 June 2015
[33] G Cheng C Yang X Yao L Guo and J Han ldquoWhenDeep Learning Meets Metric Learning Remote Sensing ImageScene Classification via Learning Discriminative CNNsrdquo IEEETransactions on Geoscience and Remote Sensing pp 1ndash11
[34] J He J Feng X Liu et al ldquoMobile product search with Bag ofHash Bits and boundary rerankingrdquo in Proceedings of the 2012IEEE Conference on Computer Vision and Pattern RecognitionCVPR 2012 pp 3005ndash3012 USA June 2012
Security and Communication Networks 13
[35] F Shen Y Mu Y Yang et al ldquoClassification by retrievalBinarizing data and classifiersrdquo in Proceedings of the 40thInternational ACM SIGIR Conference on Research and Develop-ment in Information Retrieval SIGIR 2017 pp 595ndash604 JapanAugust 2017
[36] P Li S Zhao andR Zhang ldquoA cluster analysis selection strategyfor supersaturated designsrdquo Computational Statistics amp DataAnalysis vol 54 no 6 pp 1605ndash1612 2010
[37] A Pradeep S Mridula and P Mohanan ldquoHigh securityidentity tags using spiral resonatorsrdquo Cmc-Computers Materialsamp Continua vol 52 no 3 pp 187ndash196 2016
[38] Y Cao Z Zhou X Sun and C Gao ldquoCoverless informationhiding based on the molecular structure images of materialrdquoComputers Materials and Continua vol 54 no 2 pp 197ndash2072018
[39] Y LiuH Peng and JWang ldquoVerifiable diversity ranking searchover encrypted outsourced datardquo Cmc-Computers Materials ampContinua vol 55 no 1 pp 037ndash057 2018
[40] K Lin J Lu C-S Chen and J Zhou ldquoLearning compactbinary descriptors with unsupervised deep neural networksrdquo inProceedings of the 2016 IEEEConference onComputer Vision andPattern Recognition CVPR 2016 pp 1183ndash1192 USA July 2016
[41] T Do A Doan and N Cheung ldquoLearning to Hash with BinaryDeep Neural Networkrdquo in Computer Vision ndash ECCV 2016vol 9909 of Lecture Notes in Computer Science pp 219ndash234Springer International Publishing Cham 2016
[42] Rui Zhang Di Xiao and Yanting Chang ldquoA Novel ImageAuthentication with Tamper Localization and Self-Recovery inEncrypted Domain Based on Compressive Sensingrdquo Securityand Communication Networks vol 2018 Article ID 1591206 15pages 2018
[43] Xia ShuangKui and JianbinWu ldquoAModification-Free Steganog-raphy Method Based on Image Information Entropyrdquo Securityand Communication Networks vol 2018 Article ID 6256872 8pages 2018
[44] J Zhang B Qu and N Xiu ldquoSome projection-like methods forthe generalized Nash equilibriardquo Computational Optimizationand Applications vol 45 no 1 pp 89ndash109 2010
[45] Biao Qu and Jing Zhao ldquoMethods for Solving Generalized NashEquilibriumrdquo Journal of Applied Mathematics vol 2013 ArticleID 762165 6 pages 2013
[46] CWang CMa and J Zhou ldquoA new class of exact penalty func-tions and penalty algorithmsrdquo Journal of Global Optimizationvol 58 no 1 pp 51ndash73 2014
[47] Y Wang X Sun and F Meng ldquoOn the conditional andpartial trade credit policywith capital constraints A StackelbergModelrdquo Applied Mathematical Modelling vol 40 no 1 pp 1ndash182016
[48] S Lian and Y Duan ldquoSmoothing of the lower-order exactpenalty function for inequality constrained optimizationrdquo Jour-nal of Inequalities and Applications Paper No 185 12 pages2016
Chemical EngineeringInternational Journal of Antennas and
Propagation
International Journal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Navigation and Observation
International Journal of
Hindawi
wwwhindawicom Volume 2018
Advances in
Multimedia
Submit your manuscripts atwwwhindawicom
Security and Communication Networks 9
IDLHLSHDHDeepBitUH-BDNN
16 32 48 64 96 128 2568Bits
0
01
02
03
04
05
06
07
08Av
erag
e Ret
rieva
l Pre
cisio
n
(a) Average-Retrieval Precision
IDLHLSHDHDeepBitUH-BDNN
01 02 03 04 05 06 07 08 09 100Recall
0
01
02
03
04
05
06
07
08
09
Prec
ision
(b) Precision-recall at 48 bits
Figure 4 Five kinds of algorithm retrieval performance comparison on CIFAR-10 dataset
8 bits16 bits32 bits48 bits64 bits128 bits
01 02 03 04 05 06 07 08 09 10Recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
Figure 5 Precision-recall curves with lengths
color image library Caltech 256 This is because theinformation contained in a more colorful image is morecomplex Adding a hidden layer can learn more detailsof the image and help improve the accuracy of thesearch
Next we tested the performance of secondary imageretrieval The value of k in the secondary search is 20 andthe experimental results are shown in Figure 10 As can
Level 1Level 2Level 3Level 4
01 02 03 04 05 06 07 08 09 1000Recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
Figure 6 Precision-recall curves with different code differenthidden layers
be seen from the results secondary retrieval can effectivelyimprove the retrieval accuracy when the number of codedbits is small However with the increase in the number ofencoding bits the results of the secondary search and theaccuracy of the primary search are not much different Thisis because the shorter the hash sequence is the easier thefeature vectors with different original features are mappedto the same hash code In order to make up for the errors
10 Security and Communication Networks
IDLHLSHDHDeepBitUH-BDNN
16 32 48 64 96 128 2568Bits
0
01
02
03
04
05
06
07
08
09
1Av
erag
e-Re
trie
val P
reci
sion
(a) Average-Retrieval Precision
IDLHLSHDHDeepBitUH-BDNN
01 02 03 04 05 06 07 08 09 1000Precision-recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
(b) Precision-recall at 48 bits
Figure 7 Five kinds of algorithm retrieval performance comparison on Caltech 256 set
8 bits16 bits32 bits48 bits64 bits128 bits
01 02 03 04 05 06 07 08 09 10Recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
Figure 8 Precision-recall curves with different code lengths
caused by the short hash code it is necessary to performsecondary searchWhen the number of encoding bits is smalla secondary retrieval method is used in IDLH and the searchaccuracy rate can be improved at the expense of a small searchspeed
Level 1Level 2Level 3Level 4
01 02 03 04 05 06 07 08 09 10Recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
Figure 9 Precision-recall curves with different hidden layers
5 Conclusion
With the rapid development of data storage and digital pro-cess more and more digital information is transformed andtransmitted over the Internet day by day which brings peoplea series of security problems as well as convenience [42] Theresearches on digital image security that is image encryptionimage data hiding and image authentication become moreimportant than ever [38 39] The most essential problem of
Security and Communication Networks 11
First retrievalSecond retrieval
16 32 48 64 96 128 2568Bits
07
08
09
1
Aver
age-
Retr
ieva
l Pre
cisio
n
Figure 10 Average-Retrieval Precision with first and secondretrieval
image recognition is to extract robust features The qualityof feature extraction is directly related to the effect of recog-nition so most of the previous work on image recognitionis spent on artificial design features [43] In recent yearsthe emergence of deep learning technology has changed thestatus of artificial design classification characteristics Deeplearning technology simulates the mechanism of humanvisual system information classification processing from themost primitive image pixels to lower edge features then tothe target components that are combined on the edge andfinally to the whole target depth learning can be combinedby layer by layer The high-level feature is the combination oflow level features From low level to high level features aremore and more abstract and show semantics more and moreFrom the underlying features to the combination of high-levelfeatures it is the depth of learning that is done by itself Itdoes not require manual intervention Compared with thecharacteristics of the artificial design this combination offeatures can be closer to the semantic expression
In terms of illegal image retrieval the traditional recog-nition method should establish a recognition model foreach type of recognition task In the actual application arecognition model needs a recognition server If there aremany identification tasks the cost is too high We used thedeep neural network to recognize the illegal image it onlyneeds to collect the samples of every kind of illegal imageand participate in the training of the deep neural networkFinally a multiclassification recognition model is trainedWhen classifying unknown samples deep neural networkaccounting calculates the probability that the image belongsto each class
We all know that in the image detection process theaccuracy and recall rate are mutually influential Ideally bothmust be high but in general the accuracy is high and the
recall rate is low the recall rate is high and the accuracy islow For image retrieval we need to improve the accuracyunder the condition of guaranteeing the recall rate Forimage disease surveillance and anti-illegal images we need toenhance the recall under the condition of ensuring accuracyTherefore in different application scenarios in order toachieve a balance between accuracy and recall perhaps somegame theory (such as Nash Equilibrium [44 45]) and penaltyfunction [46ndash48] can provide related optimization solutions
In this paper we proposed an improved deep-learning-hashing approach IDLH which optimized over two majorimage retrieval process
(a) In the feature extraction process the self-encodednetwork of the look-ahead type is trained by using unlabeledimage data and the expression of robust image features islearned This unlabeled learning method does not requireimage library labeling and reduces the requirements for theimage library At the same time it also takes advantage of thedeep learning networks strong learning ability and obtainsbetter image feature expression than ordinary algorithms
(b) On the index structure a secondary search is pro-posed which further increases the accuracy of the search atthe expense of very little retrieval time
Through experiments the algorithm proposed in thispaper is compared with other classic hashing algorithms onmultiple evaluation indicators Firstly we tested the learningnetworks of different code lengths and depths in order totest their effect on the retrieval system and then tested theperformance of the secondary search Through the above-mentioned series of experiments for different parameters theeffectiveness of the improved deep learning hash retrievalalgorithm proposed in this paper is verified and throughthe experimental data the good retrieval results are provedIn addition the proposed deep hashing training strategycan also be potentially applied to other hashing problemsinvolving data similarity computation
Data Availability
The data used to support the findings of this study areincluded within the article
Conflicts of Interest
The authors declare that they have no conflicts of interest
Acknowledgments
The work was funded by the National Natural ScienceFoundation of China (Grants nos 61206138 and 61373016)
References
[1] R Datta D Joshi J Li and J Z Wang ldquoImage retrieval ideasinfluences and trends of the new agerdquoACMComputing Surveysvol 40 no 2 article 5 2008
[2] G Shakhnarovich T Darrell and P Indyk Nearest-NeighborMethods in Learning andVisionTheory and PracticeMITPressCambridge MA USA 2006
12 Security and Communication Networks
[3] A Gionis P Indyk and R Motwani ldquoSimilarity search in highdimensions via hashingrdquo in 25th Int Conf pp 518ndash529 1999
[4] Z Pan J Lei Y Zhang and F L Wang ldquoAdaptive fractional-Pixel motion estimation skipped algorithm for efficient HEVCmotion estimationrdquoACMTransactions onMultimedia Comput-ing Communications and Applications (TOMM) vol 14 no 1pp 1ndash19 2018
[5] G-L Tian M Wang and L Song ldquoVariable selection in thehigh-dimensional continuous generalized linear model withcurrent status datardquo Journal of Applied Statistics vol 41 no 3pp 467ndash483 2014
[6] M Datar N Immorlica P Indyk and V S Mirrokni ldquoLocality-sensitive hashing scheme based on p-stable distributionsrdquo inProceedings of the 20th Annual Symposium on ComputationalGeometry (SCG rsquo04) pp 253ndash262 ACM June 2004
[7] B Kulis P Jain and K Grauman ldquoFast similarity search forlearned metricsrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 31 no 12 pp 2143ndash2157 2009
[8] M Raginsky and S Lazebnik ldquoLocality-sensitive binary codesfrom shift-invariant kernelsrdquo in Proceedings of the 23rd AnnualConference on Neural Information Processing Systems NIPS2009 pp 1509ndash1517 Canada December 2009
[9] L Qi X Zhang W Dou and Q Ni ldquoA distributed locality-sensitive hashing-based approach for cloud service recommen-dation from multi-source datardquo IEEE Journal on Selected Areasin Communications vol 35 no 11 pp 2616ndash2624 2017
[10] M A Carreira-Perpinan and R Raziperchikolaei ldquoHashingwith binary autoencodersrdquo in Proceedings of the IEEE Confer-ence on Computer Vision and Pattern Recognition CVPR 2015pp 557ndash566 USA June 2015
[11] J Wang S Kumar and S-F Chang ldquoSemi-supervised hashingfor large-scale searchrdquo IEEE Transactions on Pattern Analysisand Machine Intelligence vol 34 no 12 pp 2393ndash2406 2012
[12] Y Gong S Lazebnik and A Gordo ldquoIterative quantizationa Procrustean approach to learning binary codes for large-scale image retrievalrdquo in Proceedings of the IEEE Conference onComputer Vision and Pattern Recognition (CVPR rsquo11) pp 2916ndash2929 June 2011
[13] W Kong and W J Li ldquoIsotropic hashingrdquo NIPS vol 25 2012[14] M Norouzi and D J Fleet ldquoMinimal loss hashing for compact
binary codesrdquo in Proceedings of the 28th International Confer-ence on Machine Learning ICML 2011 pp 353ndash360 USA July2011
[15] J Wang W Liu A X Sun and Y-G Jiang ldquoLearning hashcodes with listwise supervisionrdquo in Proceedings of the 2013 14thIEEE International Conference on Computer Vision ICCV 2013pp 3032ndash3039 Australia December 2013
[16] G Lin C Shen Q Shi A Van Den Hengel and D Suter ldquoFastsupervised hashing with decision trees for high-dimensionaldatardquo in Proceedings of 27th IEEE Conference on ComputerVision and Pattern Recognition CVPRrsquo pp 1971ndash1978 USA2014
[17] Y Gong S Kumar H A Rowley and S Lazebnik ldquoLearningbinary codes for high-dimensional data using bilinear projec-tionsrdquo in Proceedings of the 26th IEEE Conference on ComputerVision and Pattern Recognition CVPR 2013 pp 484ndash491 USAJune 2013
[18] W Liu J Wang Y Mu and S Kumar ldquoCompact hyperplanehashing with bilinear functionsrdquo in The 29th InternationalConference on Machine Learning (ICML12) pp 467ndash474 2012
[19] Y Weiss A Torralba and R Fergus ldquoSpectral hashingrdquo inProceedings of the 22nd Annual Conference on Neural Informa-tion Processing Systems (NIPS rsquo08) pp 1753ndash1760 VancouverCanada December 2008
[20] W Liu J Wang S Kumar and S F Chang ldquoHashing withgraphsrdquo inThe 28th international conference on machine learn-ing (ICML11) 2011
[21] F Shen X Zhou Y Yang J Song H T Shen and D Tao ldquoA fastoptimization method for general binary code learningrdquo IEEETransactions on Image Processing vol 25 no 12 pp 5610ndash56212016
[22] F Shen W Liu S Zhang Y Yang and H T Shen ldquoLearningbinary codes for maximum inner product searchrdquo in Proceed-ings of the 15th IEEE International Conference on ComputerVision ICCV 2015 pp 4148ndash4156 Chile December 2015
[23] A Krizhevsky I Sutskever andG EHinton ldquoImagenet classifi-cation with deep convolutional neural networksrdquo in Proceedingsof the 26th Annual Conference on Neural Information ProcessingSystems (NIPS rsquo12) pp 1097ndash1105 Lake Tahoe Nev USADecember 2012
[24] G E Hinton and R R Salakhutdinov ldquoReducing the dimen-sionality of data with neural networksrdquo The American Associa-tion for the Advancement of Science Science vol 313 no 5786pp 504ndash507 2006
[25] A Torralba R Fergus and Y Weiss ldquoSmall codes and largeimage databases for recognitionrdquo in Proceedings of the IEEEComputer Society Conference on Computer Vision and PatternRecognition (CVPR rsquo08) pp 1ndash8 2008
[26] R Salakhutdinov andG Hinton ldquoLearning a nonlinear embed-ding by preserving class neighbourhood structurerdquo Journal ofMachine Learning Research vol 2 pp 412ndash419 2007
[27] V E Liong J Lu GWang P Moulin and J Zhou ldquoDeep hash-ing for compact binary codes learningrdquo in Proceedings of theIEEE Conference on Computer Vision and Pattern RecognitionCVPR 2015 pp 2475ndash2483 USA June 2015
[28] Y Gong S Lazebnik A Gordo and F Perronnin ldquoIterativequantization A procrustean approach to learning binary codesfor large-scale image retrievalrdquo IEEE Transactions on PatternAnalysis andMachine Intelligence vol 35 no 12 pp 2916ndash29292013
[29] W Liu J Wang R Ji Y-G Jiang and S-F Chang ldquoSupervisedhashing with kernelsrdquo in Proceedings of the IEEE Conference onComputer Vision and Pattern Recognition (CVPR rsquo12) pp 2074ndash2081 Providence RI USA June 2012
[30] J Masci A Bronstein M Bronstein and P SprechmannldquoSparse similarity-preserving hashingrdquo in Int Conf LearnRepresent pp 1ndash13 2014
[31] B Kulis and K Grauman ldquoKernelized locality-sensitive hash-ingrdquo IEEE Transactions on Pattern Analysis and Machine Intel-ligence vol 34 no 6 pp 1092ndash1104 2012
[32] F Zhao YHuang LWang and T Tan ldquoDeep semantic rankingbased hashing for multi-label image retrievalrdquo in Proceedings ofIEEE Conference on Computer Vision and Pattern RecognitionCVPR 2015 pp 1556ndash1564 June 2015
[33] G Cheng C Yang X Yao L Guo and J Han ldquoWhenDeep Learning Meets Metric Learning Remote Sensing ImageScene Classification via Learning Discriminative CNNsrdquo IEEETransactions on Geoscience and Remote Sensing pp 1ndash11
[34] J He J Feng X Liu et al ldquoMobile product search with Bag ofHash Bits and boundary rerankingrdquo in Proceedings of the 2012IEEE Conference on Computer Vision and Pattern RecognitionCVPR 2012 pp 3005ndash3012 USA June 2012
Security and Communication Networks 13
[35] F Shen Y Mu Y Yang et al ldquoClassification by retrievalBinarizing data and classifiersrdquo in Proceedings of the 40thInternational ACM SIGIR Conference on Research and Develop-ment in Information Retrieval SIGIR 2017 pp 595ndash604 JapanAugust 2017
[36] P Li S Zhao andR Zhang ldquoA cluster analysis selection strategyfor supersaturated designsrdquo Computational Statistics amp DataAnalysis vol 54 no 6 pp 1605ndash1612 2010
[37] A Pradeep S Mridula and P Mohanan ldquoHigh securityidentity tags using spiral resonatorsrdquo Cmc-Computers Materialsamp Continua vol 52 no 3 pp 187ndash196 2016
[38] Y Cao Z Zhou X Sun and C Gao ldquoCoverless informationhiding based on the molecular structure images of materialrdquoComputers Materials and Continua vol 54 no 2 pp 197ndash2072018
[39] Y LiuH Peng and JWang ldquoVerifiable diversity ranking searchover encrypted outsourced datardquo Cmc-Computers Materials ampContinua vol 55 no 1 pp 037ndash057 2018
[40] K Lin J Lu C-S Chen and J Zhou ldquoLearning compactbinary descriptors with unsupervised deep neural networksrdquo inProceedings of the 2016 IEEEConference onComputer Vision andPattern Recognition CVPR 2016 pp 1183ndash1192 USA July 2016
[41] T Do A Doan and N Cheung ldquoLearning to Hash with BinaryDeep Neural Networkrdquo in Computer Vision ndash ECCV 2016vol 9909 of Lecture Notes in Computer Science pp 219ndash234Springer International Publishing Cham 2016
[42] Rui Zhang Di Xiao and Yanting Chang ldquoA Novel ImageAuthentication with Tamper Localization and Self-Recovery inEncrypted Domain Based on Compressive Sensingrdquo Securityand Communication Networks vol 2018 Article ID 1591206 15pages 2018
[43] Xia ShuangKui and JianbinWu ldquoAModification-Free Steganog-raphy Method Based on Image Information Entropyrdquo Securityand Communication Networks vol 2018 Article ID 6256872 8pages 2018
[44] J Zhang B Qu and N Xiu ldquoSome projection-like methods forthe generalized Nash equilibriardquo Computational Optimizationand Applications vol 45 no 1 pp 89ndash109 2010
[45] Biao Qu and Jing Zhao ldquoMethods for Solving Generalized NashEquilibriumrdquo Journal of Applied Mathematics vol 2013 ArticleID 762165 6 pages 2013
[46] CWang CMa and J Zhou ldquoA new class of exact penalty func-tions and penalty algorithmsrdquo Journal of Global Optimizationvol 58 no 1 pp 51ndash73 2014
[47] Y Wang X Sun and F Meng ldquoOn the conditional andpartial trade credit policywith capital constraints A StackelbergModelrdquo Applied Mathematical Modelling vol 40 no 1 pp 1ndash182016
[48] S Lian and Y Duan ldquoSmoothing of the lower-order exactpenalty function for inequality constrained optimizationrdquo Jour-nal of Inequalities and Applications Paper No 185 12 pages2016
Chemical EngineeringInternational Journal of Antennas and
Propagation
International Journal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Navigation and Observation
International Journal of
Hindawi
wwwhindawicom Volume 2018
Advances in
Multimedia
Submit your manuscripts atwwwhindawicom
10 Security and Communication Networks
IDLHLSHDHDeepBitUH-BDNN
16 32 48 64 96 128 2568Bits
0
01
02
03
04
05
06
07
08
09
1Av
erag
e-Re
trie
val P
reci
sion
(a) Average-Retrieval Precision
IDLHLSHDHDeepBitUH-BDNN
01 02 03 04 05 06 07 08 09 1000Precision-recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
(b) Precision-recall at 48 bits
Figure 7 Five kinds of algorithm retrieval performance comparison on Caltech 256 set
8 bits16 bits32 bits48 bits64 bits128 bits
01 02 03 04 05 06 07 08 09 10Recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
Figure 8 Precision-recall curves with different code lengths
caused by the short hash code it is necessary to performsecondary searchWhen the number of encoding bits is smalla secondary retrieval method is used in IDLH and the searchaccuracy rate can be improved at the expense of a small searchspeed
Level 1Level 2Level 3Level 4
01 02 03 04 05 06 07 08 09 10Recall
0
01
02
03
04
05
06
07
08
09
1
Prec
ision
Figure 9 Precision-recall curves with different hidden layers
5 Conclusion
With the rapid development of data storage and digital pro-cess more and more digital information is transformed andtransmitted over the Internet day by day which brings peoplea series of security problems as well as convenience [42] Theresearches on digital image security that is image encryptionimage data hiding and image authentication become moreimportant than ever [38 39] The most essential problem of
Security and Communication Networks 11
First retrievalSecond retrieval
16 32 48 64 96 128 2568Bits
07
08
09
1
Aver
age-
Retr
ieva
l Pre
cisio
n
Figure 10 Average-Retrieval Precision with first and secondretrieval
image recognition is to extract robust features The qualityof feature extraction is directly related to the effect of recog-nition so most of the previous work on image recognitionis spent on artificial design features [43] In recent yearsthe emergence of deep learning technology has changed thestatus of artificial design classification characteristics Deeplearning technology simulates the mechanism of humanvisual system information classification processing from themost primitive image pixels to lower edge features then tothe target components that are combined on the edge andfinally to the whole target depth learning can be combinedby layer by layer The high-level feature is the combination oflow level features From low level to high level features aremore and more abstract and show semantics more and moreFrom the underlying features to the combination of high-levelfeatures it is the depth of learning that is done by itself Itdoes not require manual intervention Compared with thecharacteristics of the artificial design this combination offeatures can be closer to the semantic expression
In terms of illegal image retrieval the traditional recog-nition method should establish a recognition model foreach type of recognition task In the actual application arecognition model needs a recognition server If there aremany identification tasks the cost is too high We used thedeep neural network to recognize the illegal image it onlyneeds to collect the samples of every kind of illegal imageand participate in the training of the deep neural networkFinally a multiclassification recognition model is trainedWhen classifying unknown samples deep neural networkaccounting calculates the probability that the image belongsto each class
We all know that in the image detection process theaccuracy and recall rate are mutually influential Ideally bothmust be high but in general the accuracy is high and the
recall rate is low the recall rate is high and the accuracy islow For image retrieval we need to improve the accuracyunder the condition of guaranteeing the recall rate Forimage disease surveillance and anti-illegal images we need toenhance the recall under the condition of ensuring accuracyTherefore in different application scenarios in order toachieve a balance between accuracy and recall perhaps somegame theory (such as Nash Equilibrium [44 45]) and penaltyfunction [46ndash48] can provide related optimization solutions
In this paper we proposed an improved deep-learning-hashing approach IDLH which optimized over two majorimage retrieval process
(a) In the feature extraction process the self-encodednetwork of the look-ahead type is trained by using unlabeledimage data and the expression of robust image features islearned This unlabeled learning method does not requireimage library labeling and reduces the requirements for theimage library At the same time it also takes advantage of thedeep learning networks strong learning ability and obtainsbetter image feature expression than ordinary algorithms
(b) On the index structure a secondary search is pro-posed which further increases the accuracy of the search atthe expense of very little retrieval time
Through experiments the algorithm proposed in thispaper is compared with other classic hashing algorithms onmultiple evaluation indicators Firstly we tested the learningnetworks of different code lengths and depths in order totest their effect on the retrieval system and then tested theperformance of the secondary search Through the above-mentioned series of experiments for different parameters theeffectiveness of the improved deep learning hash retrievalalgorithm proposed in this paper is verified and throughthe experimental data the good retrieval results are provedIn addition the proposed deep hashing training strategycan also be potentially applied to other hashing problemsinvolving data similarity computation
Data Availability
The data used to support the findings of this study areincluded within the article
Conflicts of Interest
The authors declare that they have no conflicts of interest
Acknowledgments
The work was funded by the National Natural ScienceFoundation of China (Grants nos 61206138 and 61373016)
References
[1] R Datta D Joshi J Li and J Z Wang ldquoImage retrieval ideasinfluences and trends of the new agerdquoACMComputing Surveysvol 40 no 2 article 5 2008
[2] G Shakhnarovich T Darrell and P Indyk Nearest-NeighborMethods in Learning andVisionTheory and PracticeMITPressCambridge MA USA 2006
12 Security and Communication Networks
[3] A Gionis P Indyk and R Motwani ldquoSimilarity search in highdimensions via hashingrdquo in 25th Int Conf pp 518ndash529 1999
[4] Z Pan J Lei Y Zhang and F L Wang ldquoAdaptive fractional-Pixel motion estimation skipped algorithm for efficient HEVCmotion estimationrdquoACMTransactions onMultimedia Comput-ing Communications and Applications (TOMM) vol 14 no 1pp 1ndash19 2018
[5] G-L Tian M Wang and L Song ldquoVariable selection in thehigh-dimensional continuous generalized linear model withcurrent status datardquo Journal of Applied Statistics vol 41 no 3pp 467ndash483 2014
[6] M Datar N Immorlica P Indyk and V S Mirrokni ldquoLocality-sensitive hashing scheme based on p-stable distributionsrdquo inProceedings of the 20th Annual Symposium on ComputationalGeometry (SCG rsquo04) pp 253ndash262 ACM June 2004
[7] B Kulis P Jain and K Grauman ldquoFast similarity search forlearned metricsrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 31 no 12 pp 2143ndash2157 2009
[8] M Raginsky and S Lazebnik ldquoLocality-sensitive binary codesfrom shift-invariant kernelsrdquo in Proceedings of the 23rd AnnualConference on Neural Information Processing Systems NIPS2009 pp 1509ndash1517 Canada December 2009
[9] L Qi X Zhang W Dou and Q Ni ldquoA distributed locality-sensitive hashing-based approach for cloud service recommen-dation from multi-source datardquo IEEE Journal on Selected Areasin Communications vol 35 no 11 pp 2616ndash2624 2017
[10] M A Carreira-Perpinan and R Raziperchikolaei ldquoHashingwith binary autoencodersrdquo in Proceedings of the IEEE Confer-ence on Computer Vision and Pattern Recognition CVPR 2015pp 557ndash566 USA June 2015
[11] J Wang S Kumar and S-F Chang ldquoSemi-supervised hashingfor large-scale searchrdquo IEEE Transactions on Pattern Analysisand Machine Intelligence vol 34 no 12 pp 2393ndash2406 2012
[12] Y Gong S Lazebnik and A Gordo ldquoIterative quantizationa Procrustean approach to learning binary codes for large-scale image retrievalrdquo in Proceedings of the IEEE Conference onComputer Vision and Pattern Recognition (CVPR rsquo11) pp 2916ndash2929 June 2011
[13] W Kong and W J Li ldquoIsotropic hashingrdquo NIPS vol 25 2012[14] M Norouzi and D J Fleet ldquoMinimal loss hashing for compact
binary codesrdquo in Proceedings of the 28th International Confer-ence on Machine Learning ICML 2011 pp 353ndash360 USA July2011
[15] J Wang W Liu A X Sun and Y-G Jiang ldquoLearning hashcodes with listwise supervisionrdquo in Proceedings of the 2013 14thIEEE International Conference on Computer Vision ICCV 2013pp 3032ndash3039 Australia December 2013
[16] G Lin C Shen Q Shi A Van Den Hengel and D Suter ldquoFastsupervised hashing with decision trees for high-dimensionaldatardquo in Proceedings of 27th IEEE Conference on ComputerVision and Pattern Recognition CVPRrsquo pp 1971ndash1978 USA2014
[17] Y Gong S Kumar H A Rowley and S Lazebnik ldquoLearningbinary codes for high-dimensional data using bilinear projec-tionsrdquo in Proceedings of the 26th IEEE Conference on ComputerVision and Pattern Recognition CVPR 2013 pp 484ndash491 USAJune 2013
[18] W Liu J Wang Y Mu and S Kumar ldquoCompact hyperplanehashing with bilinear functionsrdquo in The 29th InternationalConference on Machine Learning (ICML12) pp 467ndash474 2012
[19] Y Weiss A Torralba and R Fergus ldquoSpectral hashingrdquo inProceedings of the 22nd Annual Conference on Neural Informa-tion Processing Systems (NIPS rsquo08) pp 1753ndash1760 VancouverCanada December 2008
[20] W Liu J Wang S Kumar and S F Chang ldquoHashing withgraphsrdquo inThe 28th international conference on machine learn-ing (ICML11) 2011
[21] F Shen X Zhou Y Yang J Song H T Shen and D Tao ldquoA fastoptimization method for general binary code learningrdquo IEEETransactions on Image Processing vol 25 no 12 pp 5610ndash56212016
[22] F Shen W Liu S Zhang Y Yang and H T Shen ldquoLearningbinary codes for maximum inner product searchrdquo in Proceed-ings of the 15th IEEE International Conference on ComputerVision ICCV 2015 pp 4148ndash4156 Chile December 2015
[23] A Krizhevsky I Sutskever andG EHinton ldquoImagenet classifi-cation with deep convolutional neural networksrdquo in Proceedingsof the 26th Annual Conference on Neural Information ProcessingSystems (NIPS rsquo12) pp 1097ndash1105 Lake Tahoe Nev USADecember 2012
[24] G E Hinton and R R Salakhutdinov ldquoReducing the dimen-sionality of data with neural networksrdquo The American Associa-tion for the Advancement of Science Science vol 313 no 5786pp 504ndash507 2006
[25] A Torralba R Fergus and Y Weiss ldquoSmall codes and largeimage databases for recognitionrdquo in Proceedings of the IEEEComputer Society Conference on Computer Vision and PatternRecognition (CVPR rsquo08) pp 1ndash8 2008
[26] R Salakhutdinov andG Hinton ldquoLearning a nonlinear embed-ding by preserving class neighbourhood structurerdquo Journal ofMachine Learning Research vol 2 pp 412ndash419 2007
[27] V E Liong J Lu GWang P Moulin and J Zhou ldquoDeep hash-ing for compact binary codes learningrdquo in Proceedings of theIEEE Conference on Computer Vision and Pattern RecognitionCVPR 2015 pp 2475ndash2483 USA June 2015
[28] Y Gong S Lazebnik A Gordo and F Perronnin ldquoIterativequantization A procrustean approach to learning binary codesfor large-scale image retrievalrdquo IEEE Transactions on PatternAnalysis andMachine Intelligence vol 35 no 12 pp 2916ndash29292013
[29] W Liu J Wang R Ji Y-G Jiang and S-F Chang ldquoSupervisedhashing with kernelsrdquo in Proceedings of the IEEE Conference onComputer Vision and Pattern Recognition (CVPR rsquo12) pp 2074ndash2081 Providence RI USA June 2012
[30] J Masci A Bronstein M Bronstein and P SprechmannldquoSparse similarity-preserving hashingrdquo in Int Conf LearnRepresent pp 1ndash13 2014
[31] B Kulis and K Grauman ldquoKernelized locality-sensitive hash-ingrdquo IEEE Transactions on Pattern Analysis and Machine Intel-ligence vol 34 no 6 pp 1092ndash1104 2012
[32] F Zhao YHuang LWang and T Tan ldquoDeep semantic rankingbased hashing for multi-label image retrievalrdquo in Proceedings ofIEEE Conference on Computer Vision and Pattern RecognitionCVPR 2015 pp 1556ndash1564 June 2015
[33] G Cheng C Yang X Yao L Guo and J Han ldquoWhenDeep Learning Meets Metric Learning Remote Sensing ImageScene Classification via Learning Discriminative CNNsrdquo IEEETransactions on Geoscience and Remote Sensing pp 1ndash11
[34] J He J Feng X Liu et al ldquoMobile product search with Bag ofHash Bits and boundary rerankingrdquo in Proceedings of the 2012IEEE Conference on Computer Vision and Pattern RecognitionCVPR 2012 pp 3005ndash3012 USA June 2012
Security and Communication Networks 13
[35] F Shen Y Mu Y Yang et al ldquoClassification by retrievalBinarizing data and classifiersrdquo in Proceedings of the 40thInternational ACM SIGIR Conference on Research and Develop-ment in Information Retrieval SIGIR 2017 pp 595ndash604 JapanAugust 2017
[36] P Li S Zhao andR Zhang ldquoA cluster analysis selection strategyfor supersaturated designsrdquo Computational Statistics amp DataAnalysis vol 54 no 6 pp 1605ndash1612 2010
[37] A Pradeep S Mridula and P Mohanan ldquoHigh securityidentity tags using spiral resonatorsrdquo Cmc-Computers Materialsamp Continua vol 52 no 3 pp 187ndash196 2016
[38] Y Cao Z Zhou X Sun and C Gao ldquoCoverless informationhiding based on the molecular structure images of materialrdquoComputers Materials and Continua vol 54 no 2 pp 197ndash2072018
[39] Y LiuH Peng and JWang ldquoVerifiable diversity ranking searchover encrypted outsourced datardquo Cmc-Computers Materials ampContinua vol 55 no 1 pp 037ndash057 2018
[40] K Lin J Lu C-S Chen and J Zhou ldquoLearning compactbinary descriptors with unsupervised deep neural networksrdquo inProceedings of the 2016 IEEEConference onComputer Vision andPattern Recognition CVPR 2016 pp 1183ndash1192 USA July 2016
[41] T Do A Doan and N Cheung ldquoLearning to Hash with BinaryDeep Neural Networkrdquo in Computer Vision ndash ECCV 2016vol 9909 of Lecture Notes in Computer Science pp 219ndash234Springer International Publishing Cham 2016
[42] Rui Zhang Di Xiao and Yanting Chang ldquoA Novel ImageAuthentication with Tamper Localization and Self-Recovery inEncrypted Domain Based on Compressive Sensingrdquo Securityand Communication Networks vol 2018 Article ID 1591206 15pages 2018
[43] Xia ShuangKui and JianbinWu ldquoAModification-Free Steganog-raphy Method Based on Image Information Entropyrdquo Securityand Communication Networks vol 2018 Article ID 6256872 8pages 2018
[44] J Zhang B Qu and N Xiu ldquoSome projection-like methods forthe generalized Nash equilibriardquo Computational Optimizationand Applications vol 45 no 1 pp 89ndash109 2010
[45] Biao Qu and Jing Zhao ldquoMethods for Solving Generalized NashEquilibriumrdquo Journal of Applied Mathematics vol 2013 ArticleID 762165 6 pages 2013
[46] CWang CMa and J Zhou ldquoA new class of exact penalty func-tions and penalty algorithmsrdquo Journal of Global Optimizationvol 58 no 1 pp 51ndash73 2014
[47] Y Wang X Sun and F Meng ldquoOn the conditional andpartial trade credit policywith capital constraints A StackelbergModelrdquo Applied Mathematical Modelling vol 40 no 1 pp 1ndash182016
[48] S Lian and Y Duan ldquoSmoothing of the lower-order exactpenalty function for inequality constrained optimizationrdquo Jour-nal of Inequalities and Applications Paper No 185 12 pages2016
Chemical EngineeringInternational Journal of Antennas and
Propagation
International Journal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Navigation and Observation
International Journal of
Hindawi
wwwhindawicom Volume 2018
Advances in
Multimedia
Submit your manuscripts atwwwhindawicom
Security and Communication Networks 11
First retrievalSecond retrieval
16 32 48 64 96 128 2568Bits
07
08
09
1
Aver
age-
Retr
ieva
l Pre
cisio
n
Figure 10 Average-Retrieval Precision with first and secondretrieval
image recognition is to extract robust features The qualityof feature extraction is directly related to the effect of recog-nition so most of the previous work on image recognitionis spent on artificial design features [43] In recent yearsthe emergence of deep learning technology has changed thestatus of artificial design classification characteristics Deeplearning technology simulates the mechanism of humanvisual system information classification processing from themost primitive image pixels to lower edge features then tothe target components that are combined on the edge andfinally to the whole target depth learning can be combinedby layer by layer The high-level feature is the combination oflow level features From low level to high level features aremore and more abstract and show semantics more and moreFrom the underlying features to the combination of high-levelfeatures it is the depth of learning that is done by itself Itdoes not require manual intervention Compared with thecharacteristics of the artificial design this combination offeatures can be closer to the semantic expression
In terms of illegal image retrieval the traditional recog-nition method should establish a recognition model foreach type of recognition task In the actual application arecognition model needs a recognition server If there aremany identification tasks the cost is too high We used thedeep neural network to recognize the illegal image it onlyneeds to collect the samples of every kind of illegal imageand participate in the training of the deep neural networkFinally a multiclassification recognition model is trainedWhen classifying unknown samples deep neural networkaccounting calculates the probability that the image belongsto each class
We all know that in the image detection process theaccuracy and recall rate are mutually influential Ideally bothmust be high but in general the accuracy is high and the
recall rate is low the recall rate is high and the accuracy islow For image retrieval we need to improve the accuracyunder the condition of guaranteeing the recall rate Forimage disease surveillance and anti-illegal images we need toenhance the recall under the condition of ensuring accuracyTherefore in different application scenarios in order toachieve a balance between accuracy and recall perhaps somegame theory (such as Nash Equilibrium [44 45]) and penaltyfunction [46ndash48] can provide related optimization solutions
In this paper we proposed an improved deep-learning-hashing approach IDLH which optimized over two majorimage retrieval process
(a) In the feature extraction process the self-encodednetwork of the look-ahead type is trained by using unlabeledimage data and the expression of robust image features islearned This unlabeled learning method does not requireimage library labeling and reduces the requirements for theimage library At the same time it also takes advantage of thedeep learning networks strong learning ability and obtainsbetter image feature expression than ordinary algorithms
(b) On the index structure a secondary search is pro-posed which further increases the accuracy of the search atthe expense of very little retrieval time
Through experiments the algorithm proposed in thispaper is compared with other classic hashing algorithms onmultiple evaluation indicators Firstly we tested the learningnetworks of different code lengths and depths in order totest their effect on the retrieval system and then tested theperformance of the secondary search Through the above-mentioned series of experiments for different parameters theeffectiveness of the improved deep learning hash retrievalalgorithm proposed in this paper is verified and throughthe experimental data the good retrieval results are provedIn addition the proposed deep hashing training strategycan also be potentially applied to other hashing problemsinvolving data similarity computation
Data Availability
The data used to support the findings of this study areincluded within the article
Conflicts of Interest
The authors declare that they have no conflicts of interest
Acknowledgments
The work was funded by the National Natural ScienceFoundation of China (Grants nos 61206138 and 61373016)
References
[1] R Datta D Joshi J Li and J Z Wang ldquoImage retrieval ideasinfluences and trends of the new agerdquoACMComputing Surveysvol 40 no 2 article 5 2008
[2] G Shakhnarovich T Darrell and P Indyk Nearest-NeighborMethods in Learning andVisionTheory and PracticeMITPressCambridge MA USA 2006
12 Security and Communication Networks
[3] A Gionis P Indyk and R Motwani ldquoSimilarity search in highdimensions via hashingrdquo in 25th Int Conf pp 518ndash529 1999
[4] Z Pan J Lei Y Zhang and F L Wang ldquoAdaptive fractional-Pixel motion estimation skipped algorithm for efficient HEVCmotion estimationrdquoACMTransactions onMultimedia Comput-ing Communications and Applications (TOMM) vol 14 no 1pp 1ndash19 2018
[5] G-L Tian M Wang and L Song ldquoVariable selection in thehigh-dimensional continuous generalized linear model withcurrent status datardquo Journal of Applied Statistics vol 41 no 3pp 467ndash483 2014
[6] M Datar N Immorlica P Indyk and V S Mirrokni ldquoLocality-sensitive hashing scheme based on p-stable distributionsrdquo inProceedings of the 20th Annual Symposium on ComputationalGeometry (SCG rsquo04) pp 253ndash262 ACM June 2004
[7] B Kulis P Jain and K Grauman ldquoFast similarity search forlearned metricsrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 31 no 12 pp 2143ndash2157 2009
[8] M Raginsky and S Lazebnik ldquoLocality-sensitive binary codesfrom shift-invariant kernelsrdquo in Proceedings of the 23rd AnnualConference on Neural Information Processing Systems NIPS2009 pp 1509ndash1517 Canada December 2009
[9] L Qi X Zhang W Dou and Q Ni ldquoA distributed locality-sensitive hashing-based approach for cloud service recommen-dation from multi-source datardquo IEEE Journal on Selected Areasin Communications vol 35 no 11 pp 2616ndash2624 2017
[10] M A Carreira-Perpinan and R Raziperchikolaei ldquoHashingwith binary autoencodersrdquo in Proceedings of the IEEE Confer-ence on Computer Vision and Pattern Recognition CVPR 2015pp 557ndash566 USA June 2015
[11] J Wang S Kumar and S-F Chang ldquoSemi-supervised hashingfor large-scale searchrdquo IEEE Transactions on Pattern Analysisand Machine Intelligence vol 34 no 12 pp 2393ndash2406 2012
[12] Y Gong S Lazebnik and A Gordo ldquoIterative quantizationa Procrustean approach to learning binary codes for large-scale image retrievalrdquo in Proceedings of the IEEE Conference onComputer Vision and Pattern Recognition (CVPR rsquo11) pp 2916ndash2929 June 2011
[13] W Kong and W J Li ldquoIsotropic hashingrdquo NIPS vol 25 2012[14] M Norouzi and D J Fleet ldquoMinimal loss hashing for compact
binary codesrdquo in Proceedings of the 28th International Confer-ence on Machine Learning ICML 2011 pp 353ndash360 USA July2011
[15] J Wang W Liu A X Sun and Y-G Jiang ldquoLearning hashcodes with listwise supervisionrdquo in Proceedings of the 2013 14thIEEE International Conference on Computer Vision ICCV 2013pp 3032ndash3039 Australia December 2013
[16] G Lin C Shen Q Shi A Van Den Hengel and D Suter ldquoFastsupervised hashing with decision trees for high-dimensionaldatardquo in Proceedings of 27th IEEE Conference on ComputerVision and Pattern Recognition CVPRrsquo pp 1971ndash1978 USA2014
[17] Y Gong S Kumar H A Rowley and S Lazebnik ldquoLearningbinary codes for high-dimensional data using bilinear projec-tionsrdquo in Proceedings of the 26th IEEE Conference on ComputerVision and Pattern Recognition CVPR 2013 pp 484ndash491 USAJune 2013
[18] W Liu J Wang Y Mu and S Kumar ldquoCompact hyperplanehashing with bilinear functionsrdquo in The 29th InternationalConference on Machine Learning (ICML12) pp 467ndash474 2012
[19] Y Weiss A Torralba and R Fergus ldquoSpectral hashingrdquo inProceedings of the 22nd Annual Conference on Neural Informa-tion Processing Systems (NIPS rsquo08) pp 1753ndash1760 VancouverCanada December 2008
[20] W Liu J Wang S Kumar and S F Chang ldquoHashing withgraphsrdquo inThe 28th international conference on machine learn-ing (ICML11) 2011
[21] F Shen X Zhou Y Yang J Song H T Shen and D Tao ldquoA fastoptimization method for general binary code learningrdquo IEEETransactions on Image Processing vol 25 no 12 pp 5610ndash56212016
[22] F Shen W Liu S Zhang Y Yang and H T Shen ldquoLearningbinary codes for maximum inner product searchrdquo in Proceed-ings of the 15th IEEE International Conference on ComputerVision ICCV 2015 pp 4148ndash4156 Chile December 2015
[23] A Krizhevsky I Sutskever andG EHinton ldquoImagenet classifi-cation with deep convolutional neural networksrdquo in Proceedingsof the 26th Annual Conference on Neural Information ProcessingSystems (NIPS rsquo12) pp 1097ndash1105 Lake Tahoe Nev USADecember 2012
[24] G E Hinton and R R Salakhutdinov ldquoReducing the dimen-sionality of data with neural networksrdquo The American Associa-tion for the Advancement of Science Science vol 313 no 5786pp 504ndash507 2006
[25] A Torralba R Fergus and Y Weiss ldquoSmall codes and largeimage databases for recognitionrdquo in Proceedings of the IEEEComputer Society Conference on Computer Vision and PatternRecognition (CVPR rsquo08) pp 1ndash8 2008
[26] R Salakhutdinov andG Hinton ldquoLearning a nonlinear embed-ding by preserving class neighbourhood structurerdquo Journal ofMachine Learning Research vol 2 pp 412ndash419 2007
[27] V E Liong J Lu GWang P Moulin and J Zhou ldquoDeep hash-ing for compact binary codes learningrdquo in Proceedings of theIEEE Conference on Computer Vision and Pattern RecognitionCVPR 2015 pp 2475ndash2483 USA June 2015
[28] Y Gong S Lazebnik A Gordo and F Perronnin ldquoIterativequantization A procrustean approach to learning binary codesfor large-scale image retrievalrdquo IEEE Transactions on PatternAnalysis andMachine Intelligence vol 35 no 12 pp 2916ndash29292013
[29] W Liu J Wang R Ji Y-G Jiang and S-F Chang ldquoSupervisedhashing with kernelsrdquo in Proceedings of the IEEE Conference onComputer Vision and Pattern Recognition (CVPR rsquo12) pp 2074ndash2081 Providence RI USA June 2012
[30] J Masci A Bronstein M Bronstein and P SprechmannldquoSparse similarity-preserving hashingrdquo in Int Conf LearnRepresent pp 1ndash13 2014
[31] B Kulis and K Grauman ldquoKernelized locality-sensitive hash-ingrdquo IEEE Transactions on Pattern Analysis and Machine Intel-ligence vol 34 no 6 pp 1092ndash1104 2012
[32] F Zhao YHuang LWang and T Tan ldquoDeep semantic rankingbased hashing for multi-label image retrievalrdquo in Proceedings ofIEEE Conference on Computer Vision and Pattern RecognitionCVPR 2015 pp 1556ndash1564 June 2015
[33] G Cheng C Yang X Yao L Guo and J Han ldquoWhenDeep Learning Meets Metric Learning Remote Sensing ImageScene Classification via Learning Discriminative CNNsrdquo IEEETransactions on Geoscience and Remote Sensing pp 1ndash11
[34] J He J Feng X Liu et al ldquoMobile product search with Bag ofHash Bits and boundary rerankingrdquo in Proceedings of the 2012IEEE Conference on Computer Vision and Pattern RecognitionCVPR 2012 pp 3005ndash3012 USA June 2012
Security and Communication Networks 13
[35] F Shen Y Mu Y Yang et al ldquoClassification by retrievalBinarizing data and classifiersrdquo in Proceedings of the 40thInternational ACM SIGIR Conference on Research and Develop-ment in Information Retrieval SIGIR 2017 pp 595ndash604 JapanAugust 2017
[36] P Li S Zhao andR Zhang ldquoA cluster analysis selection strategyfor supersaturated designsrdquo Computational Statistics amp DataAnalysis vol 54 no 6 pp 1605ndash1612 2010
[37] A Pradeep S Mridula and P Mohanan ldquoHigh securityidentity tags using spiral resonatorsrdquo Cmc-Computers Materialsamp Continua vol 52 no 3 pp 187ndash196 2016
[38] Y Cao Z Zhou X Sun and C Gao ldquoCoverless informationhiding based on the molecular structure images of materialrdquoComputers Materials and Continua vol 54 no 2 pp 197ndash2072018
[39] Y LiuH Peng and JWang ldquoVerifiable diversity ranking searchover encrypted outsourced datardquo Cmc-Computers Materials ampContinua vol 55 no 1 pp 037ndash057 2018
[40] K Lin J Lu C-S Chen and J Zhou ldquoLearning compactbinary descriptors with unsupervised deep neural networksrdquo inProceedings of the 2016 IEEEConference onComputer Vision andPattern Recognition CVPR 2016 pp 1183ndash1192 USA July 2016
[41] T Do A Doan and N Cheung ldquoLearning to Hash with BinaryDeep Neural Networkrdquo in Computer Vision ndash ECCV 2016vol 9909 of Lecture Notes in Computer Science pp 219ndash234Springer International Publishing Cham 2016
[42] Rui Zhang Di Xiao and Yanting Chang ldquoA Novel ImageAuthentication with Tamper Localization and Self-Recovery inEncrypted Domain Based on Compressive Sensingrdquo Securityand Communication Networks vol 2018 Article ID 1591206 15pages 2018
[43] Xia ShuangKui and JianbinWu ldquoAModification-Free Steganog-raphy Method Based on Image Information Entropyrdquo Securityand Communication Networks vol 2018 Article ID 6256872 8pages 2018
[44] J Zhang B Qu and N Xiu ldquoSome projection-like methods forthe generalized Nash equilibriardquo Computational Optimizationand Applications vol 45 no 1 pp 89ndash109 2010
[45] Biao Qu and Jing Zhao ldquoMethods for Solving Generalized NashEquilibriumrdquo Journal of Applied Mathematics vol 2013 ArticleID 762165 6 pages 2013
[46] CWang CMa and J Zhou ldquoA new class of exact penalty func-tions and penalty algorithmsrdquo Journal of Global Optimizationvol 58 no 1 pp 51ndash73 2014
[47] Y Wang X Sun and F Meng ldquoOn the conditional andpartial trade credit policywith capital constraints A StackelbergModelrdquo Applied Mathematical Modelling vol 40 no 1 pp 1ndash182016
[48] S Lian and Y Duan ldquoSmoothing of the lower-order exactpenalty function for inequality constrained optimizationrdquo Jour-nal of Inequalities and Applications Paper No 185 12 pages2016
Chemical EngineeringInternational Journal of Antennas and
Propagation
International Journal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Navigation and Observation
International Journal of
Hindawi
wwwhindawicom Volume 2018
Advances in
Multimedia
Submit your manuscripts atwwwhindawicom
12 Security and Communication Networks
[3] A Gionis P Indyk and R Motwani ldquoSimilarity search in highdimensions via hashingrdquo in 25th Int Conf pp 518ndash529 1999
[4] Z Pan J Lei Y Zhang and F L Wang ldquoAdaptive fractional-Pixel motion estimation skipped algorithm for efficient HEVCmotion estimationrdquoACMTransactions onMultimedia Comput-ing Communications and Applications (TOMM) vol 14 no 1pp 1ndash19 2018
[5] G-L Tian M Wang and L Song ldquoVariable selection in thehigh-dimensional continuous generalized linear model withcurrent status datardquo Journal of Applied Statistics vol 41 no 3pp 467ndash483 2014
[6] M Datar N Immorlica P Indyk and V S Mirrokni ldquoLocality-sensitive hashing scheme based on p-stable distributionsrdquo inProceedings of the 20th Annual Symposium on ComputationalGeometry (SCG rsquo04) pp 253ndash262 ACM June 2004
[7] B Kulis P Jain and K Grauman ldquoFast similarity search forlearned metricsrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 31 no 12 pp 2143ndash2157 2009
[8] M Raginsky and S Lazebnik ldquoLocality-sensitive binary codesfrom shift-invariant kernelsrdquo in Proceedings of the 23rd AnnualConference on Neural Information Processing Systems NIPS2009 pp 1509ndash1517 Canada December 2009
[9] L Qi X Zhang W Dou and Q Ni ldquoA distributed locality-sensitive hashing-based approach for cloud service recommen-dation from multi-source datardquo IEEE Journal on Selected Areasin Communications vol 35 no 11 pp 2616ndash2624 2017
[10] M A Carreira-Perpinan and R Raziperchikolaei ldquoHashingwith binary autoencodersrdquo in Proceedings of the IEEE Confer-ence on Computer Vision and Pattern Recognition CVPR 2015pp 557ndash566 USA June 2015
[11] J Wang S Kumar and S-F Chang ldquoSemi-supervised hashingfor large-scale searchrdquo IEEE Transactions on Pattern Analysisand Machine Intelligence vol 34 no 12 pp 2393ndash2406 2012
[12] Y Gong S Lazebnik and A Gordo ldquoIterative quantizationa Procrustean approach to learning binary codes for large-scale image retrievalrdquo in Proceedings of the IEEE Conference onComputer Vision and Pattern Recognition (CVPR rsquo11) pp 2916ndash2929 June 2011
[13] W Kong and W J Li ldquoIsotropic hashingrdquo NIPS vol 25 2012[14] M Norouzi and D J Fleet ldquoMinimal loss hashing for compact
binary codesrdquo in Proceedings of the 28th International Confer-ence on Machine Learning ICML 2011 pp 353ndash360 USA July2011
[15] J Wang W Liu A X Sun and Y-G Jiang ldquoLearning hashcodes with listwise supervisionrdquo in Proceedings of the 2013 14thIEEE International Conference on Computer Vision ICCV 2013pp 3032ndash3039 Australia December 2013
[16] G Lin C Shen Q Shi A Van Den Hengel and D Suter ldquoFastsupervised hashing with decision trees for high-dimensionaldatardquo in Proceedings of 27th IEEE Conference on ComputerVision and Pattern Recognition CVPRrsquo pp 1971ndash1978 USA2014
[17] Y Gong S Kumar H A Rowley and S Lazebnik ldquoLearningbinary codes for high-dimensional data using bilinear projec-tionsrdquo in Proceedings of the 26th IEEE Conference on ComputerVision and Pattern Recognition CVPR 2013 pp 484ndash491 USAJune 2013
[18] W Liu J Wang Y Mu and S Kumar ldquoCompact hyperplanehashing with bilinear functionsrdquo in The 29th InternationalConference on Machine Learning (ICML12) pp 467ndash474 2012
[19] Y Weiss A Torralba and R Fergus ldquoSpectral hashingrdquo inProceedings of the 22nd Annual Conference on Neural Informa-tion Processing Systems (NIPS rsquo08) pp 1753ndash1760 VancouverCanada December 2008
[20] W Liu J Wang S Kumar and S F Chang ldquoHashing withgraphsrdquo inThe 28th international conference on machine learn-ing (ICML11) 2011
[21] F Shen X Zhou Y Yang J Song H T Shen and D Tao ldquoA fastoptimization method for general binary code learningrdquo IEEETransactions on Image Processing vol 25 no 12 pp 5610ndash56212016
[22] F Shen W Liu S Zhang Y Yang and H T Shen ldquoLearningbinary codes for maximum inner product searchrdquo in Proceed-ings of the 15th IEEE International Conference on ComputerVision ICCV 2015 pp 4148ndash4156 Chile December 2015
[23] A Krizhevsky I Sutskever andG EHinton ldquoImagenet classifi-cation with deep convolutional neural networksrdquo in Proceedingsof the 26th Annual Conference on Neural Information ProcessingSystems (NIPS rsquo12) pp 1097ndash1105 Lake Tahoe Nev USADecember 2012
[24] G E Hinton and R R Salakhutdinov ldquoReducing the dimen-sionality of data with neural networksrdquo The American Associa-tion for the Advancement of Science Science vol 313 no 5786pp 504ndash507 2006
[25] A Torralba R Fergus and Y Weiss ldquoSmall codes and largeimage databases for recognitionrdquo in Proceedings of the IEEEComputer Society Conference on Computer Vision and PatternRecognition (CVPR rsquo08) pp 1ndash8 2008
[26] R Salakhutdinov andG Hinton ldquoLearning a nonlinear embed-ding by preserving class neighbourhood structurerdquo Journal ofMachine Learning Research vol 2 pp 412ndash419 2007
[27] V E Liong J Lu GWang P Moulin and J Zhou ldquoDeep hash-ing for compact binary codes learningrdquo in Proceedings of theIEEE Conference on Computer Vision and Pattern RecognitionCVPR 2015 pp 2475ndash2483 USA June 2015
[28] Y Gong S Lazebnik A Gordo and F Perronnin ldquoIterativequantization A procrustean approach to learning binary codesfor large-scale image retrievalrdquo IEEE Transactions on PatternAnalysis andMachine Intelligence vol 35 no 12 pp 2916ndash29292013
[29] W Liu J Wang R Ji Y-G Jiang and S-F Chang ldquoSupervisedhashing with kernelsrdquo in Proceedings of the IEEE Conference onComputer Vision and Pattern Recognition (CVPR rsquo12) pp 2074ndash2081 Providence RI USA June 2012
[30] J Masci A Bronstein M Bronstein and P SprechmannldquoSparse similarity-preserving hashingrdquo in Int Conf LearnRepresent pp 1ndash13 2014
[31] B Kulis and K Grauman ldquoKernelized locality-sensitive hash-ingrdquo IEEE Transactions on Pattern Analysis and Machine Intel-ligence vol 34 no 6 pp 1092ndash1104 2012
[32] F Zhao YHuang LWang and T Tan ldquoDeep semantic rankingbased hashing for multi-label image retrievalrdquo in Proceedings ofIEEE Conference on Computer Vision and Pattern RecognitionCVPR 2015 pp 1556ndash1564 June 2015
[33] G Cheng C Yang X Yao L Guo and J Han ldquoWhenDeep Learning Meets Metric Learning Remote Sensing ImageScene Classification via Learning Discriminative CNNsrdquo IEEETransactions on Geoscience and Remote Sensing pp 1ndash11
[34] J He J Feng X Liu et al ldquoMobile product search with Bag ofHash Bits and boundary rerankingrdquo in Proceedings of the 2012IEEE Conference on Computer Vision and Pattern RecognitionCVPR 2012 pp 3005ndash3012 USA June 2012
Security and Communication Networks 13
[35] F Shen Y Mu Y Yang et al ldquoClassification by retrievalBinarizing data and classifiersrdquo in Proceedings of the 40thInternational ACM SIGIR Conference on Research and Develop-ment in Information Retrieval SIGIR 2017 pp 595ndash604 JapanAugust 2017
[36] P Li S Zhao andR Zhang ldquoA cluster analysis selection strategyfor supersaturated designsrdquo Computational Statistics amp DataAnalysis vol 54 no 6 pp 1605ndash1612 2010
[37] A Pradeep S Mridula and P Mohanan ldquoHigh securityidentity tags using spiral resonatorsrdquo Cmc-Computers Materialsamp Continua vol 52 no 3 pp 187ndash196 2016
[38] Y Cao Z Zhou X Sun and C Gao ldquoCoverless informationhiding based on the molecular structure images of materialrdquoComputers Materials and Continua vol 54 no 2 pp 197ndash2072018
[39] Y LiuH Peng and JWang ldquoVerifiable diversity ranking searchover encrypted outsourced datardquo Cmc-Computers Materials ampContinua vol 55 no 1 pp 037ndash057 2018
[40] K Lin J Lu C-S Chen and J Zhou ldquoLearning compactbinary descriptors with unsupervised deep neural networksrdquo inProceedings of the 2016 IEEEConference onComputer Vision andPattern Recognition CVPR 2016 pp 1183ndash1192 USA July 2016
[41] T Do A Doan and N Cheung ldquoLearning to Hash with BinaryDeep Neural Networkrdquo in Computer Vision ndash ECCV 2016vol 9909 of Lecture Notes in Computer Science pp 219ndash234Springer International Publishing Cham 2016
[42] Rui Zhang Di Xiao and Yanting Chang ldquoA Novel ImageAuthentication with Tamper Localization and Self-Recovery inEncrypted Domain Based on Compressive Sensingrdquo Securityand Communication Networks vol 2018 Article ID 1591206 15pages 2018
[43] Xia ShuangKui and JianbinWu ldquoAModification-Free Steganog-raphy Method Based on Image Information Entropyrdquo Securityand Communication Networks vol 2018 Article ID 6256872 8pages 2018
[44] J Zhang B Qu and N Xiu ldquoSome projection-like methods forthe generalized Nash equilibriardquo Computational Optimizationand Applications vol 45 no 1 pp 89ndash109 2010
[45] Biao Qu and Jing Zhao ldquoMethods for Solving Generalized NashEquilibriumrdquo Journal of Applied Mathematics vol 2013 ArticleID 762165 6 pages 2013
[46] CWang CMa and J Zhou ldquoA new class of exact penalty func-tions and penalty algorithmsrdquo Journal of Global Optimizationvol 58 no 1 pp 51ndash73 2014
[47] Y Wang X Sun and F Meng ldquoOn the conditional andpartial trade credit policywith capital constraints A StackelbergModelrdquo Applied Mathematical Modelling vol 40 no 1 pp 1ndash182016
[48] S Lian and Y Duan ldquoSmoothing of the lower-order exactpenalty function for inequality constrained optimizationrdquo Jour-nal of Inequalities and Applications Paper No 185 12 pages2016
Chemical EngineeringInternational Journal of Antennas and
Propagation
International Journal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Navigation and Observation
International Journal of
Hindawi
wwwhindawicom Volume 2018
Advances in
Multimedia
Submit your manuscripts atwwwhindawicom
Security and Communication Networks 13
[35] F Shen Y Mu Y Yang et al ldquoClassification by retrievalBinarizing data and classifiersrdquo in Proceedings of the 40thInternational ACM SIGIR Conference on Research and Develop-ment in Information Retrieval SIGIR 2017 pp 595ndash604 JapanAugust 2017
[36] P Li S Zhao andR Zhang ldquoA cluster analysis selection strategyfor supersaturated designsrdquo Computational Statistics amp DataAnalysis vol 54 no 6 pp 1605ndash1612 2010
[37] A Pradeep S Mridula and P Mohanan ldquoHigh securityidentity tags using spiral resonatorsrdquo Cmc-Computers Materialsamp Continua vol 52 no 3 pp 187ndash196 2016
[38] Y Cao Z Zhou X Sun and C Gao ldquoCoverless informationhiding based on the molecular structure images of materialrdquoComputers Materials and Continua vol 54 no 2 pp 197ndash2072018
[39] Y LiuH Peng and JWang ldquoVerifiable diversity ranking searchover encrypted outsourced datardquo Cmc-Computers Materials ampContinua vol 55 no 1 pp 037ndash057 2018
[40] K Lin J Lu C-S Chen and J Zhou ldquoLearning compactbinary descriptors with unsupervised deep neural networksrdquo inProceedings of the 2016 IEEEConference onComputer Vision andPattern Recognition CVPR 2016 pp 1183ndash1192 USA July 2016
[41] T Do A Doan and N Cheung ldquoLearning to Hash with BinaryDeep Neural Networkrdquo in Computer Vision ndash ECCV 2016vol 9909 of Lecture Notes in Computer Science pp 219ndash234Springer International Publishing Cham 2016
[42] Rui Zhang Di Xiao and Yanting Chang ldquoA Novel ImageAuthentication with Tamper Localization and Self-Recovery inEncrypted Domain Based on Compressive Sensingrdquo Securityand Communication Networks vol 2018 Article ID 1591206 15pages 2018
[43] Xia ShuangKui and JianbinWu ldquoAModification-Free Steganog-raphy Method Based on Image Information Entropyrdquo Securityand Communication Networks vol 2018 Article ID 6256872 8pages 2018
[44] J Zhang B Qu and N Xiu ldquoSome projection-like methods forthe generalized Nash equilibriardquo Computational Optimizationand Applications vol 45 no 1 pp 89ndash109 2010
[45] Biao Qu and Jing Zhao ldquoMethods for Solving Generalized NashEquilibriumrdquo Journal of Applied Mathematics vol 2013 ArticleID 762165 6 pages 2013
[46] CWang CMa and J Zhou ldquoA new class of exact penalty func-tions and penalty algorithmsrdquo Journal of Global Optimizationvol 58 no 1 pp 51ndash73 2014
[47] Y Wang X Sun and F Meng ldquoOn the conditional andpartial trade credit policywith capital constraints A StackelbergModelrdquo Applied Mathematical Modelling vol 40 no 1 pp 1ndash182016
[48] S Lian and Y Duan ldquoSmoothing of the lower-order exactpenalty function for inequality constrained optimizationrdquo Jour-nal of Inequalities and Applications Paper No 185 12 pages2016
![Evolving Hash Functions using Genetic Algorithmsajiips.com.au/papers/V4.1/V4N1.4 - Evolving Hash Functions using... · hash function called "PKP Hash" by Peter.K.Pearson [5] that](https://static.documents.pub/doc/80x56/5e3486a76e7276290f0add90/evolving-hash-functions-using-genetic-evolving-hash-functions-using-hash.jpg)
