A Study on Discrete Wavelet Transform based Texture Feature Extraction for
Image Mining
Dr. T. Karthikeyan1, P. Manikandaprabhu2 1 Associate Professor, 2 Research Scholar,
Department of Computer Science, PSG College of Arts & Science, Coimbatore. [email protected]
Abstract
This paper proposed the discrete wavelet based texture
features in Image Mining. The Proposed methodology
uses the Discrete Wavelet Transform to reduce the size
of test images. Grey Level Co-occurrence Matrix
(GLCM) is applied for all test images of Low Level
components of level 2 decomposed images to extract
the texture feature of the images. Related images are
retrieved by using different distance measure
classifiers. The experimental result shows that the
proposed method achieves comparable retrieval
performance for correlation property of GLCM of texture feature.
Keywords - Texture, Discrete Wavelet Transform, gray
level co-occurrence matrix, Distance Measures.
1. Introduction The open spread use of digital and multimedia
knowledge, storeroom; finding and recovery of images
beginning the huge database become not easy. To
facilitate economical searching and retrieving of
pictures as of the digital collection, new software and
techniques have been emerged. The need to discover a
preferred image from a huge collection is mutual by
many skilled groups including the media persons,
drawing engineers, art historians and scholars etc.
Content Based Image Retrieval (CBIR) is compared
with text or content related advance for recover similar images from the database [24, 25].
Content Based Image Retrieval (CBIR) does not
need manual annotation for each image and is not
incomplete by the availability of lexicons as a
substitute this framework utilizes the low level features
that are natural in the images, color, shape and texture.
In CBIR, some forms of parallel between images are
computed using image futures extracted from them.
Thus, users can look for images just like query images
quickly and effectively.
Fig. 1 shows the architecture of a typical CBIR
system. For each image in the image database and its
image features are extracted and the obtained feature
space (or vector) is stored in the feature database. once
a query image comes in, its feature space are going to
be compared with those within the feature database one
by one and the similar images with the smallest feature
distance will be retrieved.
Fig.1: Image Retrieval Process
CBIR may be divided in the following stages:
• Preprocessing: The image is first processed in order to
extract the features to describe the contents. The
processing involves normalization, filtering,
segmentation and object identification. The output of
this stage could be a set of significant regions and
objects. • Feature extraction: Features such as color, shape,
texture, etc. are used to describe the content of the
image. Image features can be classified into primitives.
2. Feature Extraction For the given image database [1], features are
extracted first from individual images. The visual
features like color, shape, texture or spatial features or
Feature Extraction
Image DB Query Image
Feature
Database
Query
Features
Similarity Measures
Retrieved
Images
P Manikandaprabhu et al, Int.J.Computer Technology & Applications,Vol 5 (5),1805-1811
IJCTA | Sept-Oct 2014 Available [email protected]
1805
ISSN:2229-6093
some compressed domain features. The extracted
features are delineating the feature vectors. These
feature vectors are then stored to form image feature
database. For a given query image, we similarly extract
its features and form a feature vector. This feature
vector is matched with the already stored vectors in
image feature database.
Sometimes dimensionality reduction techniques are
employed to reduce the computations. The distance
between the feature vector of the query image and those of the images in the database is then calculated. The
distance of a query image with itself is zero if it is in
database. Then, the distances are stored in increasing
order and retrieval is performed with the help of
indexing scheme.
The feature is distinct as a role of one or more
capacity, every of which specifies some experimental
property of an object and it quantifies some significant
characteristics of the object. We classify the various
features currently employed as follows:
• General features: Function self-regulating features such as shape, color and texture. Independent of the
abstraction level, they can be advance in divided
into:
- Pixel level: Features considered at each pixel
level, e.g. location, colour.
- Local features: Features considered above the
outcome of results is subdivision of the image band
on image segmentation or edge detection.
- Global level features: Features measured over the
whole image or simply expected sub-area of an
image.
• Domain-specific level: Application reliant features like human faces, fingerprints, and conceptual
features.
These features are typically a synthesis of low-level
features for a some specific domain.
On the other hand, all features can be closely secret
into low level features and high level features. Low
level features can be extracted directly from the
original images, whereas high-level feature extraction
must be based on low level features [2].
The vital problems of content based image retrieval
system, which are: i. Image database selection, ii. Similarity measurement, iii. Performance evaluation of
the retrieval process and iv. Low-level image features
extraction.
3. Wavelet Transform Wavelet transform has a good location property in
time and frequency domain and is exactly within the
direction of transform compression idea. The discrete
wavelet transforms states to wavelet transforms that the
wavelets are disjointedly appraised. A transform which
limits a function both in space and scaling and has
some necessary properties compared to the Fourier
transform. The transform is centred on a wavelet
matrix, which can be figured more quickly than the
analogous Fourier matrix. Most notably, the DWT is
used for signal coding, where the assets of the
transform are exploited to signify a discrete signal in an
extra redundant form, often as a preconditioning for
data compression. The discrete wavelet transform has a
vast quantity of applications in Science, Computer Science, Mathematics and Engineering.
Wavelets are functions that satisfy certain
mathematical requirements and are used in representing
data or other functions. The basic awareness of the
wavelet transform is to exemplify any arbitrary signal
“X” as a superposition of a regular of such wavelets or
basis functions. These basis functions are gained from a
single photo type wavelet called the mother wavelet by
dilation (scaling) and translation (shifts). The discrete
wavelet transform for two dimensional signals can be
defined as follows.
1 21 2 1 2
1 2
1( , ,b ,b ) ,
X b Y bw a a
a aa
(1)
Where, a= a1a2
The indexes equation.(1) w (a1, a2, b1, b2) are
called wavelet coefficients of signal X and a1, a2 are
dilation & b1, b2 are translation, ψ is the transforming
function is known as mother wavelet. Low frequencies
are examined with low temporal resolution while high
frequencies with more temporal resolution. A wavelet
transform combines both low pass and high pass
filtering in spectral decomposition of signals. In case of
discrete wavelet, the image is decomposed into a discrete set of wavelet coefficients using an orthogonal
set of basic functions. These sets are divided into four
parts such as approximation, horizontal details, vertical
details and diagonal details. Discrete Wavelet transform
[3] provide substantial improvement in picture quality
at higher compression ratio.
The Embedded Zero tree Wavelet coding is a simple,
effective progressive image coding algorithm and can
be worn for both lossless and lossy compression
systems. This algorithm works well with the proposed
coding scheme because the zero tree structure is
effective in describing the significance map of the transform coefficients, as it exploits the inherent self-
similarity of the subband image over the range of
scales, and the positioning of majority of zero valued
coefficients in the higher frequency subbands. The
EZW algorithm applies Successive Approximation
Quantization in order to provide multi-precision
representation of the transformed coefficients and to
facilitate the embedded coding. The algorithm codes
(1)
P Manikandaprabhu et al, Int.J.Computer Technology & Applications,Vol 5 (5),1805-1811
IJCTA | Sept-Oct 2014 Available [email protected]
1806
ISSN:2229-6093
the transformed coefficients in decreasing order in
several scans. Each scan of the algorithm consists of
two passes: significant map encoding and refinement
pass.
The dominant pass scans the subband structure in
zigzag, right-to-left and then top-to-bottom within each
scale, before proceeding to the next higher scale of
subband structure as presented in Fig .2. For each and
every pass, a threshold (T) is chosen against which all
the coefficients are measured and encoded as one of the following four symbols,
Significant positive – If the coefficient value is
greater than threshold T
Significant negative – If the magnitude of the
coefficient value is greater than threshold T
Zero tree root – A coefficient is encoded as zero
tree root if the coefficient and all its descendants
are insignificant with respect to threshold T
Isolated zero – If the coefficient is insignificant but
some of its descendants are significant.
2 maxlog
0 2C
T (2)
where equation.(2) Cmax is the maximum coefficient
in the subband structure. The successive approximation
quantization uses a monotonically decreasing set of
thresholds and encodes the transformed coefficients as
one of the above four labels with respect to any given
threshold. For successive significant pass encoding, the
threshold is lowered as 1
2
TT KK
and only those
coefficients not yet found to be significant in the
previous pass are scanned for encoding, and the process
is repeated until the threshold reaches zero, and results
in complete encoded bit streams.
Fig.2: EZW subband structure scanning order
In the embedded zero tree wavelet coding strategy,
developed by Shapiro, a wavelet/subband decomposition of the image is performed. The wavelet
coefficients/pixels are then grouped into Spatial
Orientation Trees. The magnitude of each wavelet
coefficients/pixels in a tree, starting with the root of the
tree, is then compared to a particular threshold T. If the
magnitude of all the wavelet coefficients/pixels in the
tree are smaller than T, the entire tree structure (that is
the root and all its descendant nodes) is coded by one
symbol, the zerotree symbol ZTR. If however, there
exit significant wavelet coefficients/pixels, then the tree
root is coded as being significant or insignificant, if its
magnitude is larger than or smaller than T, respectively.
The descendant nodes are then each examined in turn to
determine whether each is the root of a possible sub zero tree structure, or not. This process is carried out
such that all the nodes in all the trees are examined for
possible sub zero tree structures.
The significant wavelet coefficients/pixels in a tree
are coded by one of two symbols, POS or NEG,
depending on whether their actual values are positive or
negative, respectively. The process of classifying the
pixels as being ZTR, IZ, POS, or NEG is referred to as
the dominant pass in [4]. This is then followed by the
subordinate pass in which the significant wavelet
coefficients/pixels in the image are refined by determining whether their magnitudes lie within the
intervals (T, 3T/2) and (3T/2,2T). Those wavelet
coefficients/pixels whose magnitudes lie in the interval
(T, 3T/2) are represented by a 0 (LOW), whereas those
with magnitudes lying in the interval (3T/2,2T) are
represented by a 1 (HIGH). Subsequent to the
completion of both the dominant and subordinate
passes, the threshold value T is reduced by a factor of
2, and the entire process repeated. This coding strategy,
consisting of the dominant and subordinate passes
followed by the reduction in the threshold value, is
iterated until a target bit rate is achieved. The root node of each tree is located at the highest
level of the decomposition pyramid, and all its
descendants are located in different spatial frequency
bands at the same pyramid level, or clustered in groups
of 2 X 2 at lower levels of the decomposition pyramid.
An EZW decoder reconstructs the image by
progressively updating the values of each wavelet
coefficient/pixel in a tree as it receives the data. The
decoder's decisions are always synchronized to those of
the encoder.
4. Texture Features Among totally different visual characteristics like
color and shape for the analysis of various types of
images, texture is reported to be outstanding and very
important low level feature [5, 6]. Even though no
standard definition exists for texture, Sklansky [7]
outlined the texture collection of native properties among the image region with a continuing, slowly
varied or about periodic pattern. Texture gives the
information on structural arrangement of surfaces and
P Manikandaprabhu et al, Int.J.Computer Technology & Applications,Vol 5 (5),1805-1811
IJCTA | Sept-Oct 2014 Available [email protected]
1807
ISSN:2229-6093
objects on the image. Texture is not defined for a
separate pixel; it depends on the distribution of
intensity over the image. Texture possesses regularity
and scalability properties; it is represented by main
directions, contrast and sharpness. It is measured using
its distinct properties like periodicity, coarseness,
directionality and pattern complexity for efficient
image retrieval particularly on the aspects of orientation
and scale [8].
Tuceryan and Jain [9] divided the different methods for feature extraction into four main
categories, namely: structural, statistical, model-based
and transform domain. Basically, texture representation
methods can be classified into two categories: structural
and statistical. Statistical methods, including Fourier
power spectra, co-occurrence matrices, shift-invariant
principal component analysis (SPCA), Tamura
features, Wold decomposition[10], Markov random
field[11], fractal model[12] and multi-resolution
filtering techniques such as Gabor[13] and wavelet
transform[14], characterize texture by the statistical distribution of the image intensity. D. A. Clausi et. al
[15] designed the fusion texture feature with Gabor
filter and co occurrence probabilities for texture
segmentation and demonstrated that it outperforms well
for noisy images and the high dimensional feature
vector. The DWT based color cooccurrence feature for
texture classification is explained in [16].
Haralick et. al [17] proposed the methods for
representing texture features of images was grey level
co-occurrence matrices (GLCM). Haralick et. al [17]
also suggested 14 descriptors including the contrast, correlation, entropy and others. Each descriptor shows
one texture property. Therefore, many works for
example as described in [18], are devoted to selecting
those statistical descriptors derived from the co-
occurrence matrices that describe texture within the
best approach.
In [19], firstly, transforming color space from RGB
model to HSI model and then extracting color
histogram to form color feature vector. Secondly,
extracting the texture feature by using gray co-
occurrence matrix. The texture of image is an
illustration of spatial relationship of gray level image. Co-occurrence matrix is make it up based on the point
of reference and distance between image pixels. The
co-occurrence matrix C(i, j) counts the co-occurrence
of pixels with gray values i and j at a given distance d.
The distance d is outlined in polar coordinates (d, ),
with discrete length and orientation. In practice,
takes the values 0◦; 45◦; 90◦; 135◦; 180◦; 225◦; 270◦;
and 315◦. The cooccurrence matrix C(i, j) can now be
defined as follows:
1 1 2 2
1 1 2 2
2 2 1 1
(( , ), ( , )) ( ) ( )
for ( , ) , ( , )( , ) card
( , ) ( , ) ( cos , sin );
for 0<i, j<
x y x y XY XY
f x y i f x y jC i j
x y x y d d
(3)
where card {.} denotes the number of elements in the
set. Let G be the number of gray-values in the image, then the dimension of the co-occurrence matrix C (i, j)
will be N ×N. So, the computational complexity of the
co-occurrence matrix depends quadratically on the
number of gray-scales used for quantization.
A. Wavelet-Based Texture Representation
In wavelet based texture Representations, a
specific feature of this method is representation and
analysis of signals in different scales, i.e., under
different resolutions. The image is described by a
hierarchical structure each level of which represents the original signal with a certain degree of detail.
Tamura et al. [20] presented an approach to
describing texture on the basis on human visual
perception. They suggested coarseness, contrast,
directionality, line-likeness, regularity and roughness
equivalent to the six texture properties that were
recognized as visually significant in the course of
psychological experiments. Howarth and Ruger [18,
21] noticed that the parameters describing the primary
three properties coarseness, contrast and directionality
are rather effective in classifying and searching images by texture. The set of all points for one image is
referred to as Tamura image.
Texture analysis by means of the Gabor filters is a
special case of the wavelet approach. This is the most
frequently used method in image retrieval by texture. In
most of the CBIR systems primarily based in Gabor
wavelet [22, 23], the mean and standard deviation of
the distribution of the wavelet transform coefficients
are used to construct the feature vector.
B. Correlation property
Correlation property shows the linear dependency
of gray level values in the co-occurrence matrix. It
presents how a reference pixel is related to its
neighbour, 0 is uncorrelated, 1 is perfectly correlated.
P Manikandaprabhu et al, Int.J.Computer Technology & Applications,Vol 5 (5),1805-1811
IJCTA | Sept-Oct 2014 Available [email protected]
1808
ISSN:2229-6093
i ij
i j
i j
i
i j
j
j i
2
i i
i j
2
j j
i j
(ij)C(i,j)-μ μ
Correlation= (4)σ σ
where
μ = i C(i,j)
μ = j C(i,j)
σ = (i-μ ) C(i,j)
σ = (j-μ ) C(i,j)
5. Distance Measures Distance metrics are considered among the enquiry
image and each image in the database. This procedure
is frequent awaiting all the images in the file have been
related with the query image. Remoteness among two images is used to searching the similarities between
query image and the images in the database. Distance
measures like the city block, Standard Euclidean
distance method include used to found the comparison
of feature vectors. In this paper, we use the Euclidean
distance, Standard Euclidean distance and also city
block distance are used to compare the similarity
between the images.
A. City-Block distance (L1)
It computes the distance that may be go to get from
one point to other data point. The amount of the dissimilarity of their corresponding example.
1
n
i i
i
d x y
(5)
B. Euclidean distance (L2)
Euclidean distance is nearly everyone often used to
evaluate profiles of respondents diagonally variables.
This is the nearly all commonly-used metric distance
measure. Euclidean distance is the rectangle root of the
amount of the squared differences between equivalent elements of the two vectors.
2
1
( )n
i i
i
d x y
(6)
C. Standard Euclidean Distance (Std L2)
Standardized Euclidean distance earnings Euclidean distance is planned on regular facts.
Standardized value = (Original value - mean)/Standard
Deviation
1
n
i i
i
d x y
(7)
6. Performance Measures Assessment of retrieval presentation is a critical
trouble in content-based image retrieval (CBIR). Many
different methods for measuring the performance of a
system have been created and used by researchers. The
most common evaluation measures used in CBIR are
precision and recall which are defined as,
Number of relevant images retrievedPrecision =
Total number of images retrieved
7. Experimental Results
Corel image database of 1000 images have been
used. Each image is of size 256x384. There are 10
classes in this database like Africans, Buildings, Buses,
Dinosaurs, Elephants, Flowers, Mountains and Peoples
in database. Each class contains 100 images. The
retrieval efficiency and effectiveness of the proposed
texture feature and Distance measures are experimented
with the popular image database Corel image database and the experimental results are presented in this
section. This experiment gives the comparison of
performance measures of CBIR for the metric City
Block distance (L1), Euclidean distance (L2) and
Standard Euclidean Distance (Std L2). Here we
compared the GLCM Correlation property of precision
of Class names Buses, Dinosaurs, Elephants,
Mountains and peoples.
A. Graph Results
The graph results in Fig 3. Shows performance analysis
related to retrieval accuracy of various class name.
Fig. 3: Precision of Correlation in Each Class
P Manikandaprabhu et al, Int.J.Computer Technology & Applications,Vol 5 (5),1805-1811
IJCTA | Sept-Oct 2014 Available [email protected]
1809
ISSN:2229-6093
The graph results in Fig 4. Shows the average retrieval
accuracy of various distance measures.
Fig. 4: Average of Precision Efficiency
TABLE I
Detailed Precision of Correlation by Class Name
Class Name L1 L2 Std L2
Buses 94 96 92
Dinosaur 96 96 94
Elephant 88 90 84
Mountain 86 86 80
People 70 76 70
Average 86.8 88.8 84
VII. CONCLUSION In this paper, discrete wavelet based texture
features, associated with different distance measures have been evaluated in Corel data sets. The efficiency
and performance of the proposed system are measured
using average precision of three different distance
measures. Performance analysis comparison of
Correlation with different distance classifier therein one
Euclidean distance gives best performance than city
block and Standard Euclidean distance.
REFERENCES [1] S. Patil and S. Talbar, “Content Based Image Retrieval Using Various Distance Metrics”, Data Engineering and Management, Lecture Notes in Computer Science, Berlin Heidelberg: Springer, pp 154-161, 2012, vol. 6411.
[2] E. Saber, A.M. Tekalp, ”Integration of color, edge and texture features for automatic region-based image annotation
and retrieval”, Journal of Electronic Imaging 7(3), pp. 684–700, 1998.
[3] R. Krishnamoorthy, K. Rajavijayalakshmi and R. Punidha, "Low Complexity Hybrid Lossy To Lossless Image Coder With Combined Orthogonal Polynomials Transform And Integer Wavelet Transform”, ICTACT Journal On Image And Video Processing, Vol. 2, No. 04, pp.410-416, May 2012.
[4] Julien Reichel, Gloria Menegaz, Marcus J. Nadenau, and Murat Kunt, “Integer Wavelet Transform for Embedded Lossy to Lossless Image Compression”, IEEE Trans. Image Processing, Vol. 10, No. 3, pp. 383-392, 2001.
[5] K. Jalaja, C. Bhagvati, B. L. Deekshatulu and Arun K. Pujari, “Texture Element Feature Characterizations for CBIR”, in IEEE Proc. IGARSS '05, Vol. 2, pp. 733 - 736, 2005.
[6] T. Sikora, “The MPEG-7 visual standard for content description – an overview”, IEEE Trans. Circuits Systems and Video Technology, Vol. 11, no. 6, pp.696 – 702, 2001.
[7] J. Sklansky, “Image segmentation and feature extraction”, IEEE Trans. Systems, Man and Cybernetic, Vol.8, no. 4, pp. 237-247, 1978.
[8] H. Tamura, S. Mori and T. Yamawaki, “Texture features corresponding to visual perception”, IEEE Trans. Systems, Man and Cybernetics, Vol. 6, no. 4, pp. 460 - 473, 1976.
[9] M. Tuceryan and A. K. Jain, “Texture analysis”, In the Handbook of Pattern Recognition and Computer Vision, 207-248, 1998.
[10] F. Liu and R. Picard, “Periodicity, directionality and randomness: Wold features for image modeling and retrieval”, IEEE Trans. Pattern Analysis and Machine Intelligence, Vol. 18, no. 7, pp. 722 - 733, 1996.
[11] G. Cross and A. Jain, “Markov random field texture
models”, IEEE Trans. Pattern Analysis and Machine Intelligence, Vol. 5, no.1, pp. 25 - 39, 1983.
[12] L. M. Kaplan et al, “Fast texture database retrieval using extended fractal features”, in Storage and Retrieval for Image and Video Databases VI (Sethi, I K and Jain, R C, eds), Proc SPIE 3312, 1998, pp. 162-173.
[13] T. Chang and C.C.J. Kuo, “Texture analysis and classification with tree structured wavelet transform”, IEEE
Trans. Image Processing, Vol. 2, no. 4, pp. 173 - 188, 1992.
[14] D.S. Zhang., A. Wong., M. Indrawan, and G. Lu, “Content-based image retrieval using gabor texture features”, In Proc. of IEEE PCM’00, pp 392–395, 2000.
[15] D. A. Clausi and H. Deng, “Design Based Texture Feature Fusion Using Gabor Filters and Co- Occurrence Probabilities”, IEEE Trans. Image Processing, Vol.14, No. 7, 2005.
P Manikandaprabhu et al, Int.J.Computer Technology & Applications,Vol 5 (5),1805-1811
IJCTA | Sept-Oct 2014 Available [email protected]
1810
ISSN:2229-6093
[16] S. Arivazhagan, L. Ganesan and V. Angayakanni, “Color Texture Classification using Wavelet transform”, in Proc. of
ICCIMA’05, pp. 315-320, 2005.
[17] R.M. Haralick, K. Shanmugam, and I. Dienstein., “Textural Features for Image Classification”, IEEE Trans. Systems, Man Cybernetics., vol. 3, no. 6, pp. 610– 621, 1973.
[18] P. Howarth and S.Ruger, “Evaluation of Texture Features for Content-based Image Retrieval”, in Proc. of CIVR'04, 2004, pp. 326–334.
[19] Jiayin Kang and Wenjuan Zhang, “A Framework for
Image Retrieval with Hybrid Features”, in Proc. of CCDC, 2012, pp. 1326 – 1330.
[20] Tamura, H., Mori, S., and Yamawaki, T., “Textural Features Corresponding to Visual Perception”, IEEE Trans. Systems, Man Cybernetics, vol. 8, pp. 460–472, 1978.
[21] Howarth, P. and Ruger, S., “Robust Texture Features for Still Image Retrieval”, IEEE Proc. Vision, Image Signal Processing, vol. 152, no. 6, pp. 868–874, 2005.
[22] N. Sebe, and M.S. Lew., “Wavelet Based Texture Classification”, in IEEE Proc. of Int. Conf. on Pattern Recognition’, vol. 3, pp. 959–962, 2000.
[23] B.S. Manjunath, et. al, "Color and texture descriptors", IEEE Trans. Circuits and Systems for Video Technology, Vol.11(6), pp. 703-715, 2001.
[24] Michael S. Lew, Nicu Sebe, Chabane Djeraba and Ramesh Jain, “Content based Multimedia Information Retrieval in State of the Art and Challenge”, ACM
Multimedia Computing, Communications and Applications, Vol. 2, No. 1, pp. 1–19, Feb. 2006.
[25] T. Karthikeyan, P. Manikandaprabhu, S. Nithya, “A Survey on Text and Content Based Image Retrieval System for Image Mining”, International Journal of Engineering Research & Technology, Vol. 3 Issue 3, pp. 509-512, Mar. 2014.
P Manikandaprabhu et al, Int.J.Computer Technology & Applications,Vol 5 (5),1805-1811
IJCTA | Sept-Oct 2014 Available [email protected]
1811
ISSN:2229-6093