Automatic License Plate Recognition (ALPR) has wide range of commercial applications such as findingstolen cars, controlling access to car parks and gathering traffic flow statistics. Existing Libyan LicensePlate Recognition (LLPR) methods are not presented promising results due to their inefficient features forthe extracted characters and numbers. In this work, an improved LLPR method is presented. The method iscomposed of five stages: pre-processing, license plate extraction, character and numbers segmentation,feature extraction and license plate recognition. In the pre-processing, undesired data, such as backgroundnoises are removed. Then, the license plate is extracted using few mathematical morphologies, ConnectedComponent Analysis (CCA) and Region of Interest (ROI) extraction. After that, characters and numbersfrom the image regions of the license plate are extracted. A combination of geometrical features and Gaborfeatures are considered to represent each of the character and word in the plates. Then, the recognition isdone by using a template matching and a Probabilistic Neural Network (PNN) classification. Theperformance of the proposed method is evaluated and tested using 100 self-collected images of Libyannational license plates. The experimental results have shown that the proposed method has producedpromising results and superior than other existing methods.
Automatic vehicle identification is an imageprocessing technique of identifying vehicles bytheir license plates. The aim of using automaticvehicle identification systems is to effective trafficcontrol and security applications such as accesscontrol to Prohibited areas and tracking of wantedvehicles.
Each country has its own license plate (LP)numbering system with consideration ofcharacteristics such as: colors, language ofcharacters and style (font) and sizes includingdifference from state to state in terms of types ofLPs. Beside that, there are countries, which do notyet have an automatic license plate recognition(ALPR) system. Apart from that, the environmentaland physical conditions of the LP are also affectedthe in overall processes. Because of these reasons,research on the LP detection and recognitionprocess is still taking place. The Libyan LPs make
use of Arabic words with numbers written inEnglish compared to LP of Egypt, Saudi Arabia orother Arab countries where numbers are alsowritten in Arabic (in Libyan) as shown in figure 1.
(a) (b)
Figure 1: Libyan License Plates.
The ALPR is basically composed of imagecapturing and pre-processing, license plateextraction, characters segmentation, featureextraction and license plate recognition [1]. In the
Journal of Theoretical and Applied Information Technology15th December 2016. Vol.94. No.1
first stage, the image of a car is obtained using acamera. In this stage, it is important use the correctcamera particularly with respect to its parameters,type, resolution, shutter speed and orientation. Inthe second stage, the license plate is extracted fromthe image. Important features to be sought in thisstage include boundary, color and characters’existence. The third stage entails the segmentationof the license plate. The characters are thenextracted by projecting their color information,labeling them, or matching their positions withtemplates. The fourth stage entails recognition ofthe extracted characters. Here, classifiers ortemplate matching are employed, for instance,fuzzy classifiers and neural networks.
In this paper, automatic license platerecognition is studied for Libyan license plates.This paper is structured as follows: related worksare discussed in section 2. Section 3 presents theproposed framework for this study. Theexperimental results and performance evaluationare discussed in section 4. Finally, section 5concludes the paper.
2. RELATED WORKS
In this section, previous studies of licenseplate recognition methods are reviewed. Thisreview involves ALPR for Arabic and non-Arabiccountries.
A live and robust method for license plates’localization and identification was proposed by [2].These methods finds the optimal adaptivethreshold after the image’s intensity values ismodified to locate the vehicle’s edges. Analgorithm is then applied. This algorithm employsthe morphological operators in constructing theregions of the candidate. Each region’scharacteristics then undergoes extraction process toallow the differentiation between the license plateregion and other candidate probable regions. Thisdifferentiation is determined by the percentageanalysis of the plate’s rectangularity. A colourfilter was employed to strengthen the algorithm forthe license plate localization (LPL). The proposedalgorithm is also useful in handling distortedimages because it can effectively provideidentification and modification to the platerotation.
[3] introduced the template matchingapproach in recognizing character image. Theauthor used the method on the Egyptian and SaudiArabian systems, and in fact, this method can be
extended to other countries as well. Using thismethod, a table which consists of names of thecountries with matching Arabic characters isconstructed. The items on the table become theinputs which would then be matched with thelicense plate. A total of 400 license plate imagestaken in outdoor setting were used for this method.This method attained 90% recognition accuracyand the time used by this method to recognisevehicle plate was 1.6 seconds.
A license plate recognition system wasintroduced by [4]. This system is particularly forSaudi Arabia LPs with two templates which areutilised for recognizing Arabic and Indian numberswith limited Arabic and Latin alphabets. Pre-processing is the first step which involvesconversion of the original image to grey scale,removal of noise, detection and thresholding ofedge, and localization of the LP. Then, characterssegmentation is performed. Number anddistribution of black pixels from the horizontalprojection records in allocated regions of thecharacter image are used in feature extraction.Recognition process is based on regions and thislimits comparisons.
Abulghasem proposed an integrated methodbased on Radial Basis Function Neural Networkfor detecting and recognizing Libyan license plates[5]. The proposed method includes license platedetection, license plate extraction, character andnumbers segmentation, feature extraction andrecognition. In the detection stage, connectedcomponent analysis is used to locate uniqueobjects, from which the unwanted objects areremoved using the filtering process. Geometric andGlobal features are used to prepare the identifiedobjects before their classification as Plate and non-Plate using RBFNN. In the recognition process, forcharacter segmentation, a simple template isderived to extract and differentiate digits andArabic words. Statistical and structural features areused in feature extraction, while the classificationis performed using RBFNN. As their experimentalresults shown, the proposed method achieved 93%and 91% for accuracy rates for detection andrecognition respectively.
[6] proposed a technique for license platesegmentation. The authors employed three sets offeature vectors alongside template matching in theformation of two key modules which are licenseplate localization module and LPR module. Morethan 238 vehicle images taken from diverse sceneswith differing fonts and settings from two Arab
Journal of Theoretical and Applied Information Technology15th December 2016. Vol.94. No.1
countries were used to test this method. Asindicated by the outcomes, their proposed methodswere able to segment and extracted the charactersand words from the plates accurately, however, themethod is sensitive to noises.
[7] proposed an Iranian vehicle license platerecognition system. This system was modified toenable it to reflect the local context together with ahybrid classifier that could recognize thecharacters in the license plate. The method adoptsthe technique of modified template-matching byanalysing the target color pixels in detecting thevehicle's license plate’s location. Additionally,using a modified strip search, the authors couldlocalize the standard color-geometric templateemployed in Iran and in several other countries inEurope.
[8] proposed an algorithm for automaticlicense plate recognition. However, the algorithmproposed focuses on the Lebanese license plateswhose certain features well manipulated todecrease errors in recognition. As shown theirsimulation, there was a reduction in recognitionerrors when the Lebanese license plates are writtenin two formats, Arabic and Hindi. In order toimprove the performance, the proposed algorithmalso includes an option to benefit from thepresence of the license plates in both the front andback of the car. Nonetheless, the proposed methodwas not applied in real setting and therefore, therewere no real results for the license platerecognition.
3. PROPOSED FRAMEWORK
This section presents the proposed frameworkfor Libyan license plate recognition in this study.The proposed framework is consisted of four stagesas: Pre-processing, License plate localization,Character and number (C&N) segmentation, andLicense plate recognition as shown in figure 2.
Image Capturing andPre-processing
Character and Number (C&N)Segmentation
Feature Extraction
License Plate Recognition
Licensee Plate Extraction
Figure 2: The Proposed Framework.
3.1 Pre-processing
The pre-processing aims to get the input imageand improve the quality of the image for furtherimage processing. The pre-processing is consistedof image color mode conversion and imagebinarization.
3.1.1 Image color mode identificationThe input image is necessary in colour mode
identification. For the images, the colour mode canbe RGB or gray scale. However, the image’soriginal input is normally of RGB mode; RGBmode is colour mode. Identification of colour modecan be done using colour channel extraction. As anexample, three channels means RGB image.
3.1.2 Image binarizationBinary images comprise images with pixels
with two possible values of intensity. These pixelsare usually displayed as black and white, whereblack is usually represented by “0” and white byeither “1” or “255.” The production of binaryimages is usually by thresholding a colour orgrayscale image. Thresholding separates an objectin the image from the background. Object’s colour,which is normally white, is called the foregroundcolour; while the rest, which is normally black, iscalled the background colour.
3.2 License Plate Localization
In order to enable the extraction of numberand characters’ region in the plate, it is necessarythat the license plate is localized in the image. Inthis study, the plate localization is composed ofthree steps as: morphological operation, Connected
Journal of Theoretical and Applied Information Technology15th December 2016. Vol.94. No.1
Component Analysis (CCA) and region of interestplate extraction.
3.2.1 Morphological operation
In the context of mathematical morphology,the structuring element can be equivalent to theconvolution kernel in linear filter theory. Themorphological operation are basically involveddilation, erosion, closing and opening operators.Mathematical operations are basically required twoelements to do the operations. The mathematicaloperations are as, dilation, erosion, opening andclosing. Thus, with two sets, A and B,Minkowski addition expressed is as below:⊕ B = ⋃ ( + )∈ (1)
Meanwhile, Minkowski subtraction is expressed asbelow: ⊝ B = ∩ ∈ ( − ) (2)
Based on these two Minkowski operations, thebasic operations of mathematical morphologyof dilation and erosion are demonstrated as below:
Dilation, ( , ) = ⊕ B =⋃ ( + )∈ (3)
Erosion, ( , ) = ⊝ B =∩ ∈ ( − ) (4)
where −B = (−β|β ∈ B). The closing and openingcan be defined using the combinationof dilation and erosion. In MATLAB, theimplementation of Morphology operation is madepossible with the structuring elements of strel andbwmorph functions.
3.2.2 Connected components analysis
Connected components labelling is used toscan an image. Then, the image’s pixels aregrouped into components in accordance to theconnectivity of pixel. In other words, each pixel ina connected component share identical values ofpixel intensity and each pixel is inter-connected insome way. Using connected component labelling,an image is scanned pixel-by-pixel for theidentification of connected pixel regions.Connected pixel regions entail adjacent pixelsregions that possess identical intensity values set.
3.2.3 Region of interest extraction
In the context of this study, the area of interestregion is the license plate region based. Cropping isdone in horizontal and vertical directions of theimage. A threshold value the region of interest isidentified. Then, comparison is made between otherarea values of regions and that of the threshold sothat regions with value more than that of thresholdcan be selected. Then, based on the comparison, theregion of interest is identified. Further, employinglabelling function, the region’s coordination pointswill be detected. The points of coordination includeboth the region’s start and end point. Using thecoordination points as reference, a set of staring andend point of the region of interest is extracted andthe process of cropping is done utilising the points.
Journal of Theoretical and Applied Information Technology15th December 2016. Vol.94. No.1
In character segmentation procedure, thecharacters and numbers from the image oflicense plate are extracted. There are a variety ofaspects that make up the complex charactersegmentation task. These include frame andorientation of plat, variance of light, image noiseand space mark. Thus, in order to overcomethese difficulties in character segmentation, anumber of procedures have been proposed. Thesegmented region was analysis for differentregion properties.
3.3.1 Thresholding
Threshing is performed using athresholding function. This function comprisesim2bw which has grey-scale and a thresholdvalue. Grey-scaling converts the image of theregion to grey colour. As for the thresholdingvalue for this implementation, it is set to 0.4.
3.3.2 Labelling analysisComponent Connected Analysis (CCA) is
employed in the segmentation of characters andnumber. Following the thresholding process,extraction of characters and numbers (from the
plate) is performed using labelling analysis andgeometrical features such as area.
3.3.3 Spatial normalization
This step comprises the normalisation ofthe extracted character and number with respectto region size. Normalization is necessary due tothe difference in size of the extracted region.Thus, the resizing function, which is part ofnormalization, is applied on the extractedcharacters and numbers. Additionally,convolution filtering is employed in the spatialfrequency characteristics’ modification. Asindicated by the shape parameter, the function ofconv2 returns one subsection of the two-dimensional convolution, after which, the imageundergoes resizing using imresize function.figure 4 shows the character and wordsegmentation using thresholding, CCA andspatial normalization steps.
Journal of Theoretical and Applied Information Technology15th December 2016. Vol.94. No.1
The primary use of shape features is fordesorption of object. In this study, the features ofgeometrical shape are utilized in the featureextraction part. In MATLAB, a connectedcomponent analysis which takes the role as theregionpropos function enables the extraction offeatures of the geometrical shape. There are
numerous features for shape properties, and thesefeatures are combined as representative to thecharacter and numbers. In this study, geometricfeatures are considered to represent thecharacters and numbers. These features are as:
Meanwhile, Gabor feature is forrepresenting the characters and numbersextracted. This feature integrates withconvolution matrix used for extraction of feature.For Gabor feature, the mean value and standarddeviation of image matrix are studied. Thegeneration of the Gabor feature involves three
functions: mean value of image matrix, meanvalue of region matrix, and standard deviation.
Finally, the extracted features are necessaryin the integration and storing in a vector. Thisfeature vector could include both the geometricfeatures, which comprise 9 features, and Gaborfeature. For vector generation of this feature,
Journal of Theoretical and Applied Information Technology15th December 2016. Vol.94. No.1
simply these features by a vector refer as shownin figure 5.
Mean value ofimage matrix
Mean value ofregion matrix
Standard deviation
Gabor feature+
Geometrical features
+
Area
Centroid
...
+
Feature vector
Figure 5: Feature Vector Generation.
3.5 License Plate Recognition
This step encompasses the recognition ofthe extracted characters and numbers by way oftemplate matching and a classification such asthe techniques of probabilistic Neural network(PNN). These techniques are described below:
3.5.1 Template matching
This step aims to recognize the charactersand number in the plate. A template matchingalgorithm used in this study for this recognitionpurpose. However, before template matchingusage, it is required to have a predefinedtemplate to perform matching process betweennew plate and predefined template. In order togenerate a predefined template, we use thepreprocessing and license plate detection stepsfrom pervious phases. The results of previoussteps are segmented the character and numbersregions. Then the features for these regions areextracted, and the characters and numbers arelabeled using a manually assigned label. Aspresented in Table 1, it shows examples fordigits 1, 5 and Libya (لیبیا) word with extractedfeature and assigned label. The same processperforms for all the number 0 to 9 and Libyaword to generate the predefined samples. After
generating a predefined template, the pre-processing and license plate detection processesare performed for any new image. The matchingprocess is check whether the new image issimilar to template or not based on Euclideandistance, if it is similar the recognition processwas identifying the character or word. In thistemplate matching process, the Euclideandistance is applied for each feature of templateimage and new plate image.
3.5.2 Probabilistic neural network (PNN)As indicated, template matching is
performed to recognize the character or wordfrom the plates. However, in a situation wherethe character or word cannot be recognized, theengine of neural network will be employed topredict the segmented regions that belong to thecharacter and word’s classes. Thus, PNNpredicts the segmented region is belonged tocharacter or number classes. It also predicts thecharacter’s digit. Following the new characterprediction, the features of the character aremoved to template matching module. Thisupdates the template to allow future recognition.
The implementation of PNN necessitatestraining and testing to sets. The ensuing sections
Journal of Theoretical and Applied Information Technology15th December 2016. Vol.94. No.1
entail the description of sets’ training and testingand the PNN express usage for ourimplementation.
a) PNN Training
The input for PNN comprises extractedregions which include plates, character andnumbers regions. However, predefined datasetneeds to be used before the employment of PNNfor training. In this study, the training datasetcontains 70 images.
b) PNN Testing
PNN also requires the previouslymentioned image pre-processing steps to be doneto the image and the detection of white areas thatsatisfy the value of threshold. Then, the image’sidentified objects which are to be tested areforwarded to PNN for further testing. Asindicated by the trained data set, the outcomes
produced by the PNN are identical to that oflicense plate in the new image. The detected isgenerated from the region which was identifiedfrom the preceding step. The following providesthe details of Implanted PNN.
4. EXPERIMENTAL RESULTS ANDPERFORMANCE EVALUATION
This section presents the experimentalresults and performance evaluation for theproposed method in this study.
4.1 Experimental Setup
In this section, the details of our platformfor the implementation, experiment andperformance evaluation is presented. Table 2shows the details of experimental setup in ourplatform.
Table 2: Details Of Experimental Setup In The Proposed PlatformName Used platform
Implementation and Experimental environment MATLAB 2013a
Toolbox Image Processing and Computer Vision toolbox in MATLAB
Operating system Windows 7 64 bitCPU Intel Core 2 Quad Core 2.83 GHzRAM 4 GB MEMORY
4.2 DATASET
In this study standard license plate imagesare used for the experimental and performanceevaluation purposes. Since there is no standarddataset for the Libyan licenses images theresearchers are generally collected the imagesfrom internet. In this study, the Libyan licenseimages are also collected from internet. Thenumber of images as dataset in this experiment is100. This dataset contains standard Libyanlicense images with diffident plate format andsize. Fig shows some cars with Libyan licensesplates from dataset.
Journal of Theoretical and Applied Information Technology15th December 2016. Vol.94. No.1
4.3 PERFORMANCE EVALUATIONIn this section, the performance of proposed
method is evaluated. The performance evaluationis carried out using standard performanceevaluation metrics. The metrics are as: precision,
recall and f-score. These metrics are used tocalculate number of false and true recognition ofplates. In order to calculate the metrics somevariables are required to define as shown inTable 3.
Table 3: Definitions Of Performance Evaluation Variables.
Variables Definitions
True Positive (TP) Correctly identified characters and numbers.False Positive (FP) Incorrectly identified characters and numbers.False Negative (FN) Incorrectly identified other regions as characters or numbers.
Using the defined variables the metricscan be calculated by following equations,
= + (1)
= + (2)
= (1 + ). .( . ) + (3)
Where , , , and denotes the truepositive, false positive, false negative, precisionand recall for plate. Table 4 is the overallperformance measurement computed by theweighted harmonic of precision and recall calledas . If = 1, the called as score. Table 4shows the performance measurement of differentnumber of features using the proposed method inthis study.
Journal of Theoretical and Applied Information Technology15th December 2016. Vol.94. No.1
As depicted in Table 4, the PNN classificationmethod is tested with different number of features.The best result was obtained using all the 9 featuresincluding (These 9 features are listed in Table 1).Therefore, the optimal feature representation wasobtained based on integration of all the features.
Finally, the performance of the proposedmethod compares with other existing method toshow the outperformed method. The other existingmethod is (Abulgasem, 2012). To fair comparison,Abulgasem’s features and classification areimplemented. The features and classification ofAbulgasem used horizontal and vertical lines asfeatures and Radial Basis Classification (RBF) forclassification, character and number recognition.The features investigated in this study andAbulgasem’s features are individually implementedwith PNN and RBF classification methods. Table 5and 6 show the calculated performancemeasurements based on PNN and RBFclassification methods respectively.
Table 5. Performance measurement based on PNNFeature
extractionprocess
Recall(%)
Precision(%)
F1-score(%)
The proposedfeature
84.19 81.18 82.66
Abulgasem(2012)
81.16 78.84 79.99
Table 6: Performance Measurement Based On RBF.Feature
extractionprocess
Recall (%) Precision(%)
F1-score(%)
The proposedfeature
82.64 79.68 81.13
Abulgasem(2012)
79.66 77.39 78.51
As shown the in Tables 5 and 6, the featureextraction of this study in PNN and RBFclassification methods presented better performancein compared to Abulgasem’s feature extractionprocess. Therefore, the proposed method of thisstudy presented outperformed performance incompared to Abulgasem’s work in Libyan licenseplate recognition, because the proposed featurescontain important information for a betterrecognition.
5. CONCLUSION
An automatic Libyan license plate recognitionsystem is proposed in this study. The current studyproposed a framework which is composed of fivestages as pre-processing, license plate extraction,character and numbers segmentation, featureextraction and license plate recognition. The pre-processing step is crucial as it improves the qualityof data image to allow for visual perception orcomputational processing. In pre-processing,undesired data are removed, and throughbackground noise. License plate extraction isrequired to extract the plate while the plate islocated in image with many other objects.Mathematical morphology, Connected ComponentAnalysis (CCA) and region of interest extractionare used for plate extraction. The extracted plate issegmented to extract the characters and wordregions from the plate. In character segmentationprocedure, the characters and numbers from theimage of license plate are extracted. Thegeometrical features integrate with Gabor featuresare considered to represent the characters and wordin the plates. Using the combination of the featuresthe plates are recognised using template matchingand PNN classification techniques. The proposedframework is experimented on 100 images sampleand the performance is evaluated using standardmetrics. The performance is also compared with
Journal of Theoretical and Applied Information Technology15th December 2016. Vol.94. No.1
other existing methods to show the superior methodin license plate recognition. For the future work,this study can be extended to test the proposedmethod in real-time license plate recognition.
REFERENCES
[1] Ganoun, A., “Automatic Localization Systemof Libyan License Plates”. Vehicle Power andPropulsion Conference (VPPC), 2015 IEEE,pp. 1-4.
[2] Rastegar, S., R. Ghaderi, G. Ardeshipr & N.Asadi, “An intelligent control system using anefficient License Plate Location andRecognition Approach”, InternationalJournal of Image Processing (IJIP) Volume(3)(5), 2009,pp: 252-264.
[3] Khalil, M., “Car plate recognition using thetemplate matching method”, InternationalJournal of Computer Theory and Engineering2(5), 2010, pp: 1793-8201.
[4] Alginahi, Y. M., “Automatic arabic licenseplate recognition”, International Journal ofComputer and Electrical Engineering 3(3),2011, pp: 454-460.
[5] A Abulgasem, N., "Libyan vehicle licenseplate detection and recognition using radialbasis function", Thesis Universiti TeknologiMalaysia, Faculty of Computer Science andInformation System, 2012.
[6] Mohammad, K., S. Agaian & H. Saleh,"Arabic License Plate Recognition System",2013.
[7] Ashtari, A. H., M. J. Nordin & M. Fathy, "AnIranian License Plate Recognition SystemBased on Color Features", IntelligentTransportation Systems, IEEE Transactionson 15(4) 2014, pp: 1690-1705.
[8] El Khatib, I., Y. Sweidan, S. M. Omar & A.Al Ghouwayel, "An efficient algorithm forautomatic recognition of the Lebanese carlicense plate", Technological Advances inElectrical, Electronics and ComputerEngineering (TAEECE), 2015 ThirdInternational Conference on. pp. 185-189.
