COLOR TEXTURE CLASSIFICATION USING WAVELET TRANSFORM AND NEURAL NETWORK ENSEMBLES

Article: An auto way to categorize image
[Show abstract] [Hide abstract]
ABSTRACT: Auto categorizing is very important in computer image recognition technology. Texture is one of the keys to resolve it. Similar to human's rational intelligence is one of way to approach auto category. In this paper, we describe the common texture by analysis and term in categories theory. It means categories are useful for textures and image recognition. The combination of categories and neural network is worth to research which will gradually reveal the unity way of human's percept and rationality.01/2011;  SourceAvailable from: Sangwook Lee
Conference Paper: Wavelet domain detection of rust in steel bridge images
[Show abstract] [Hide abstract]
ABSTRACT: This paper describes the detection of rust defects on highway steel bridges, which are one of the most commonly observed defects on coating surfaces and thus have to be taken care of appropriately since they severely affect the structural integrity of bridges. A rust defect assessment method is presented that automatically detects the percentage of rust in a given digital image of bridge surface taken from a conventional digital camera. A training and detection algorithm is implemented to classify a given block of the image as rust or nonrust. The results of the algorithm are analyzed for its efficiency and possible optimization techniques are suggested.Acoustics, Speech and Signal Processing (ICASSP), 2011 IEEE International Conference on; 06/2011 · 4.63 Impact Factor  SourceAvailable from: Aksam Iftikhar[Show abstract] [Hide abstract]
ABSTRACT: In recent years, classification of colon biopsy images has become an active research area. Traditionally, colon cancer is diagnosed using microscopic analysis. However, the process is subjective and leads to considerable inter/intra observer variation. Therefore, reliable computeraided colon cancer detection techniques are in high demand. In this paper, we propose a colon biopsy image classification system, called CBIC, which benefits from discriminatory capabilities of information rich hybrid feature spaces, and performance enhancement based on ensemble classification methodology. Normal and malignant colon biopsy images differ with each other in terms of the color distribution of different biological constituents. The colors of different constituents are sharp in normal images, whereas the colors diffuse with each other in malignant images. In order to exploit this variation, two feature types, namely color components based statistical moments (CCSM) and Haralick features have been proposed, which are color components based variants of their traditional counterparts. Moreover, in normal colon biopsy images, epithelial cells possess sharp and welldefined edges. Histogram of oriented gradients (HOG) based features have been employed to exploit this information. Different combinations of hybrid features have been constructed from HOG, CCSM, and Haralick features. The minimum Redundancy Maximum Relevance (mRMR) feature selection method has been employed to select meaningful features from individual and hybrid feature sets. Finally, an ensemble classifier based on majority voting has been proposed, which classifies colon biopsy images using the selected features. Linear, RBF, and sigmoid SVM have been employed as base classifiers. The proposed system has been tested on 174 colon biopsy images, and improved performance (=98.85%) has been observed compared to previously reported studies. Additionally, the use of mRMR method has been justified by comparing the performance of CBIC on original and reduced feature sets.Computers in biology and medicine 01/2014; 47C:7692. · 1.27 Impact Factor
Page 1
Abdulkadir Sengur
October 2009 The Arabian Journal for Science and Engineering, Volume 34, Number 2B
491
COLOR TEXTURE CLASSIFICATION USING WAVELET
TRANSFORM AND NEURAL NETWORK ENSEMBLES
Abdulkadir Sengur
Firat University, Department of Electronics and Computer Science, 23119, Elazig, Turkey
ﺔـﺻﻼﺨﻟا:
ﺪﻘﻟ
ﺮﺼﺘﻘﺗ
ﻰﻠﻋ
مﺪﺨﺘﺳا
نﻵا ﻰﺘﺣ
ٌ ﺮﻴﺜآ ﻲ?ﻓ تﺎ?ﺠﻳﻮﻤﻟا ﺔ?ﻴﻨﻘﺗ ﻦﻴﺜﺣﺎ?ﺒﻟا ﻦ?ﻣ
ﻂﻘﻓ ﺔﻳدﺎﻣﺮﻟا رﻮﺼﻟا.
ﺔﻴﺒﺼﻌﻟا تﺎﻜﺒﺸﻟاو تﺎﺠﻳﻮﻤﻟا ماﺪﺨﺘﺳا.
ﻒﻴﻨﺼ?ﺗ رﻮﺼ?ﻟا ﻲ?ﻓ ﺔﺠﺴ?ﻧﻷا ﺔﻌﺠﺸ?ﻣ ﺞﺋﺎ?ﺘﻧ ﻊ?ﻣ
ﻲ?ﻓ ةﺪ?ﻳﺪﺟ ﺔ?ﻘﻳﺮﻃ حﺮﺘﻘﻨﺴ?ﻓ ،ﺚ?ﺤﺒﻟا اﺬ?ه ﻲ?ﻓ ﺎﻣأ رﻮ?ﺻ ﻦ?ﻣ ﺔﺠﺴ?ﻧﻷا ﻒﻴﻨﺼ?ﺗ
. ﻞ?ﻌﻟو ﺐﻴﻟﺎ?ﺳﻷا ﺮ?ﺜآأ ﻞ?ﻴﻠﺤﺗ ﻲ?ﻓ ً ﺎ?ﺣﺎﺠﻧ
. ﻟا ﺪ?ﻤﺘﻌﺗو ﺔ?ﻘﻳﺮﻄ
ﺔﺠﺴ?ﻧﻷا
ﻞﻴﻠﺤﺗ ﻰﻠﻋ ﺔ?ﻧﻮﻠﻣ
ﻮﻜﺘﺗ نةﺪﻳﺪﺠﻟا ﺔﻘﻳﺮﻄﻟا ﻦﻣ جاﺮﺨﺘﺳاﻲﺠﻳﻮﻤﻟا ﻞﻴﻠﺤﺘﻟا ﻦﻣ ﺔﺻﺎﺧ تﺎﻤﺳ ، ﻒﻴﻨﺼﺗ ﻢﺛ ﻩﺬ?ه
ﻲﺠﻳﻮ?ﻤﻟا . ﺖ?ﻳﺮﺟأ ﺪ?ﻗو ﻊ?ﻣ ﺔ?ﻔﻠﺘﺨﻣ برﺎ?ﺠﺗ
تزوﺎﺠﺗ ﻒﻴﻨﺼﺘﻟا ﺔﻗد 98.%
ﺔﻴﺒﺼ?ﻌﻟا تﺎﻜﺒﺸ?ﻟا ماﺪﺨﺘ?ﺳﺎﺑ تﺎﻤﺴ?ﻟا
ةدﺪ?ﻌﺘﻣ تﺎﺤ?ﺷﺮﻣ
. ﺪ?ﻳﺪﺤﺘﻟا ﻞﻴﺒ?ﺳ ﻰ?ﻠﻋ و
ﺔ?ﻴﻟﺎﻌﻓ زاﺮ?ﺑﻹ تﺎ?ﺠﻳﻮﻤﻠﻟ ﻞ?ﻴﻠﺤﺘﻟا ﻦ?ﻣ ﺎ?ﻴﺑوﺮﺗﻷاو ﺔ?ﻗﺎﻄﻟا ﻞ?ﺜﻤﺗ ﻲﺘﻟا تﺎﻤﺴﻟا ﻞﻤﻌﺘﺴﻧ فﻮﺴﻓ
ﺚﻴﺣ ﺔﻘﻳﺮﻄﻟا ﺔﻧﺎﺘﻣ ﺞﺋﺎﺘﻨﻟا تﺮﻬﻇأونإ
ﺔﺣﺮﺘﻘﻤﻟا ﺔﻴﻨﻘﺘﻟا،
Corresponding Author:
Email: ksengur@firat.edu.tr
Paper Received September 24, 2008; Paper Revised December 27, 2008; Paper Accepted February 9, 2009
Page 2
Abdulkadir Sengur
The Arabian Journal for Science and Engineering, Volume 34, Number 2B October 2009
492
ABSTRACT
The wavelet domain features have been intensively used for texture classification and texture segmentation with
encouraging results. More of the proposed multiresolution texture analysis methods are quite successful, but all the
applications of the texture analysis so far are limited to gray scale images. This paper investigates the usage of
wavelet transform and neural network ensembles for color texture classification problem. The proposed scheme is
composed of a wavelet domain feature extractor and ensembles of neural networks classifier. Entropy and energy
features are integrated to the wavelet domain feature extractor. Various experiments have been carried out with
different wavelet filters. The performed experimental studies show the efficacy of the proposed structure for color
texture classification. The highest success rate is over 98%. Moreover, we compare our results with wavelet energy
correlation signatures [2].
Key words: wavelet decomposition, neural network ensembles, texture classification, feature extraction, entropy,
energy correlation
Page 3
Abdulkadir Sengur
October 2009 The Arabian Journal for Science and Engineering, Volume 34, Number 2B
493
COLOR TEXTURE CLASSIFICATION USING WAVELET TRANSFORM AND
NEURAL NETWORK ENSEMBLES
1. INTRODUCTION
Texture analysis plays an important role in many image processing tasks, ranging from remote sensing to medical
image processing, computer vision applications, and natural scenes. A number of texture analysis methods have been
proposed in the past decades (e.g., [1]) but most of them use gray scale images, which represent the amount of
visible light at the pixel’s position, while ignoring the color information. The performance of such methods can be
improved by adding the color information because, besides texture, color is the most important property, especially
when dealing with real world images [2].
Texture can be defined as a local statistical pattern of texture primitives in an observer’s domain of interest.
Texture classification aims to assign texture labels to unknown textures, according to training samples and
classification rules. Two major issues are critical for texture classification: the texture classification algorithms and
texture feature extraction. The main aim of texture classification is to find a best matched category for a given
texture among existing textures. Texture has been analyzed extensively and many texture recognition schemes have
been proposed [24]. The important property all have in common is that they constitute an appropriate model for
relationships between the adjacent pixels of a neighborhood.
A number of researchers have proposed algorithms for texture analysis, but all the applications of the texture
analysis so far are limited to gray scale images [5]. Recently, however, both color and texture information have been
used by several researchers [2,6]. A method has been proposed by Caelli and Reye [7]. They extract features from
three spectral channels by using three multiscale isotropic filters. Van de Wouver et al. [2] proposed wavelet
energycorrelation signatures and derive the transformation of these signatures upon linear color space
transformation. The wavelet domain based cooccurrence matrix method and the second order statistical features
were used for color texture classification in the study of Arivazhagan et al. [8]. A set of features are derived and
color texture classification is done for different combinations of the features and for different color models. Karkanis
et al. [9] proposed a new approach for the detection of tumors in colonoscopic video. The proposed algorithm is
based on a color feature extraction scheme to represent the different regions in the frame sequence. The scheme is
built on the wavelet transform. The features called color wavelet covariance are based on the covariance of second
order textural measurement. A linear discriminant analysis is used for classification of the image regions. Sengur
proposed wavelet transform and ANFIS for color textures classification [6]. Crouse et al. has proposed a framework
for statistical signal modeling based on the wavelet domain hidden Markov tree [10]. The algorithm provided an
efficient approach to modeling of wavelet coefficients that are often found in real world images. Xu et al., have
shown that the wavelet coefficients have certain interdependences between color planes [11]. They used wavelet
domain hidden Markov model for color texture analysis. The proposed approach is used for modeling the
dependences between color planes as well as the interactions across scales. The wavelet coefficients at the same
location scale and subband, but with different color planes, are grouped into one vector and a multivariate Gaussian
mixture model is employed for approximating the marginal distribution of the wavelet coefficient vectors in one
scale.
In this paper, a color texture image classification scheme is proposed which uses wavelet transform and
ensembles of neural networks. Feature extraction in the wavelet domain and classification with ensembles of neural
networks are proposed. Wavelet entropies and wavelet energies of each color plane at different scales are used for
forming the feature vectors. An ensemblesbased method enables an increase in generalization performance by
combining several individual neural network trains on the same task. The experimental studies show the efficiency
of the proposed system. Moreover, a comparison of the proposed schema with the wavelet energy correlation
signatures is conducted.
The organization of this paper is as follows: in Section 2, we summarize the theory for wavelet transform, neural
networks, the ensemblesbased method, and the extraction of the energy correlation signatures; several brief
definitions are given in this section; in Section 3, the methodology and the implementation of the proposed process is
given; in Section 4, an experimental study is introduced and the classification results are shown; in Section 5, we
finally conclude the study.
2. BACKGROUND
In this section, the theoretical foundations for the expert system used in the present study are given in the
following subsections.
Page 4
Abdulkadir Sengur
The Arabian Journal for Science and Engineering, Volume 34, Number 2B October 2009
494
2.1. Discrete Wavelet Transform
Wavelet transforms are finding intense use in fields as diverse as telecommunications and biology. Because of
their suitability for analyzing nonstationary signals, they have become a powerful alternative to Fourier methods in
many medical applications, where such signals abound [12]. The main advantage of wavelets is that they have a
varying window size, being wide for slow frequencies and narrow for the fast ones, thus leading to an optimal time
frequency resolution in all the frequency ranges. Furthermore, owing to the fact that windows are adapted to the
transients of each scale, wavelets are not stationary.
The (continuous) wavelet transform of a 1D signal f (x) is defined as:
*
( , )
a b
1/2
−
( , )
a b
(W f )( , )a b , ( , )a b
ψ
( )( ) x dx
()
f f x
xb
a
a
ψ
ψ
ψψ
+∞
∫
−∞
==
−
=
(1)
where a is the scaling factor, b is the translation parameter related to the location of the window, and ψ*(x) is the
transforming function. The extension to 2D is usually performed by using a product of 1D filters. The transform is
computed by applying a filter bank to the image. The rows and columns of an image are processed separately and
down sampled by a factor of 2 in each direction, resulting in one low pass image LL and three detail images HL, LH,
and HH. Figure 1(a) shows the onelevel decomposition. The LH channel contains image information of low
horizontal frequency and high vertical frequency, the HL channel contains high horizontal frequency and low
vertical frequency, and the HH channel contains high horizontal and high vertical frequencies. The frequency
decomposition is shown in Figure 1(b). Note that in multiscale wavelet decomposition, only the LL sub band is
successively decomposed.
LL1LH1
HL1HH1
(a) 1level decomposition.
HH2
LH1
HL1HH1
LH2
HL2
LL2
(b) 2level decomposition.
Figure 1. Wavelet frequency decomposition
2.2. Artificial Neural Networks (ANN)
Artificial neural networks were originally developed by researchers who were trying to mimic the
neurophysiology of the human brain [13]. By combining many simple computing elements (neurons or units) into a
highly interconnected system, a complex phenomenon such as intelligence is produced. A schematic diagram for an
artificial neuron model is shown in Figure 2. Nowadays, neural network researchers have incorporated methods from
statistics and numerical analysis into their networks. More specifically, feedforward neural networks are a class of
flexible nonlinear regression, discriminant, and data reduction models. By detecting complex nonlinear
relationships in data, neural networks can help to make predictions about realworld problems.
Page 5
Abdulkadir Sengur
October 2009 The Arabian Journal for Science and Engineering, Volume 34, Number 2B
495
f(.) a(.)
x1
x2
xm
wi2
wi1
wim
weights
θi
outputs
yi
bias
inputs
Figure 2. Artificial neuron model [14]
The neural network node provides a variety of feedforward networks that are commonly called back
propagation networks. Backpropagation refers to the method for computing the error gradient for a feedforward
network, a straightforward application of the chain rule of elementary calculus [14,15]. By extension, back
propagation refers to various training methods that use backpropagation to compute the gradient. By further
extension, a backpropagation network is a feedforward network trained by any of various gradientdescent
techniques. There are numerous algorithms available for training neural network models; most of them can be
viewed as a straightforward application of optimization theory and statistical estimation. Most of the algorithms used
in training artificial neural networks are employing some form of gradient descent. This is done by simply taking the
derivative of the cost function with respect to the network parameters and then changing those parameters in a
gradientrelated direction. The most popular of them is the back propagation algorithm, which has different variants.
Standard back propagation is a gradient descent algorithm. It is very difficult to know which training algorithm will
be the fastest for a given problem, and the best one is usually chosen by trial and error. An ANN with a back
propagation algorithm learns by changing the connection weights, and these changes are stored as knowledge.
Neural network’s are trained by experience; when an unknown input is applied to the network it can generalize
from past experiences and product a new result. The output of the neuron net is given by Equation 2:
⎛
=+
∑
=
1j
⎟⎟
⎠
⎞
⎜⎜
⎝
−
m
ij ij
θ
(t)xwa 1) y(t
∑
=
j
−=∆
m
ij ijii
xw netf
1
θ
(2)
where X=(X1, X2… Xm ) represents the m inputs applied to the neuron, Wi represents the weights for input Xi, θi is a
bias value, and a(.) is an activation function. There are many kinds of activation function. Usually, nonlinear
activation functions such as sigmoid or step are used.
2.3. Neural Network Ensembles (NNensembles)
An ensemble of classifiers is a collection of several classifiers whose individual decisions are combined in some
way to classify the test examples [16]. It is known that an ensemble often shows much better performance than the
individual classifiers that make it up. Assume that there is an ensemble of n classifiers: {
test data x. If all the classifiers are identical, they are wrong at the same data, where an ensemble will show the same
performance as individual classifiers. However, if classifiers are different and their errors are uncorrelated, then
when
( )x is wrong, most of other classifiers except for
}
12
,,...,
n
f ff
and consider a
if( )x may be correct. Then, the result of majority voting
1
2
nn
kn k
pp
==
if
can be correct. More precisely, if the error of individual classifier is
p <
and the errors are independent, then the
probability
E
p that the result of majority voting is incorrect is
()
/ 2/2
1
2
(1)(( ) )
k
knkn
−
−<
∑∑
. When the size of
classifier n is large, the probability
E
p becomes very small.
Several methods for constructing an ensemble of classifiers have been developed in the last two decades. The
most important thing in constructing the NN ensemble is that each individual NN becomes different from the other
NNs as much as possible. This requirement can be met by using different training sets for different NNs. Some
methods for selecting the training samples are bagging, boosting, randomization, stacking, and dagging [16–8].
Among them, we put focus on bagging.
Page 6
Abdulkadir Sengur
The Arabian Journal for Science and Engineering, Volume 34, Number 2B October 2009
496
2.3.1. Bagging
In bagging, each classifier is calibrated on a randomly drawn training set with the probability of drawing any
given example being equal [18]. Samples are drawn with replacement, so that some examples may be selected
multiple times while others may not be selected at all. As a result, each classifier could return a higher test set error
than a classifier using all of the data. However, when these classifiers are combined by voting, the resulting
ensemble produces lower test set error than a single classifier. The diversity among individual classifiers
compensates for the increase in error rate of any individual classifier and improves prediction performance.
2.4. Energy Correlation Signatures
Van de Wouver et al. [2] defined the wavelet energy correlation signatures as follows:
,
, , , ,
B j l i
, , ,
B j l i
00
1
jj
mnmn
N
∑∑
N
X
B j
XXX
il
jj
Cww
N x N
==
=
(3)
where
, , ,
B j l i
m
X
w
is the wavelet coefficient at (l, i) location at j scale in B (
{
B, j
C m,n
energy signatures. They capture the energy distribution of the wavelet coefficients over the scale, sub band, and
color space for m = n and the others (m ≠ n) represent the covariance between different color spaces.
{
}
}
,,B LH HL HH
∈
) sub band and Xm is the
}
1,2,3,...,J
is called the wavelet
color space (m = 1,2,3). The set
{
mn
X ,X
1,2,3;mn;B LH,HL,HH ; j
=≤∈=
3. METHODOLOGY
The proposed color texture classification algorithm is illustrated in Figure 3. The steps involved in color texture
classification are as follows:
Step 1: The input to the NNensemble based color texture classification system is the color textures of size 512x512
[21]. We make sub images of size 128x128 by randomly choosing from the original input texture. Thus, color texture
sub images may overlap. For gray scale texture analysis, the color information is discharged by RGB to gray scale
transformation. Then the red, green, and blue components are decomposed from the color texture images and saved
for subsequent processing.
Step 2: This step involves both feature extraction and classification. The feature extraction is composed of two
layers. These are the wavelet decomposition layer and the entropy and energy calculation layer.
1. Wavelet decomposition layer: For wavelet decomposition of each of the red, green, and blue components of
color textures, the pyramid wavelet structure is used. We obtain onelevel wavelet decomposition, and save
only the three detail images HH, LH, and HL where H and L stand for the high pass and low pass band in
each of the horizontal and vertical orientations for the subsequent calculation of entropy and energy
quantities.
2. Entropy and energy calculation layer: This layer is responsible for calculating the entropy and energy
quantities of each LH, HL, and HH of the red, green, and blue components of the color texture images. Thus,
entropy and the energy quantities are the features that characterize the color texture images.
3.1. Entropy
Entropy is a quantity that is widely used in information theory and is based on probability theory [15,18].
Entropy is a common concept in many fields, mainly in mechanics, image processing, and signal processing. The
general form of the entropy is given by
2
1
() log
n
ii
i
H Xpp
=
= −∑
(4)
where X is a random variable which can be one of the values
0
, then
2
0log 0 is defined as 0. Thus, H(X) can be interpreted as representing the amount of uncertainty that
exists in the value of X. In information theory, entropy value is considered to be an average amount of information
received when the value of X is observed. In this paper, we use the norm entropy. The norm entropy H is defined as
follows:
12
,,...,
n
x xx
with probability
12
,,...,
n
p pp . Note that
if
ip =
,, , ,
B l i j
00
mm
p
NN
X
B l
X
ij
Hw
==
=∑∑
, for ()
12p
≤<
(5)
Page 7
Abdulkadir Sengur
October 2009 The Arabian Journal for Science and Engineering, Volume 34, Number 2B
497
where
color space (m = 1,2,3) [20].
, , ,
B l i j
m
X
w
is the wavelet coefficient at (i ,j) location at l scale in B (
{}
,,B LH HL HH
∈
) sub band and Xm is the
3.2. Energy
Energy is commonly used for texture analysis. In this study, we use the averaged l2norm, which is defined as
follows:
2
, , , ,
B l i j
00
1
*
()
mm
NN
X
B l
X
ij
Ew
NN
==
=
∑∑
(6)
where
color space (m = 1,2,3) [20].
, , ,
B l i j
m
X
w
is the wavelet coefficient at (i ,j) location at l scale in B (
{}
,,BLH HL HH
∈
) sub band and Xm is the
3.3. NNEnsembles:
The ensemble size is taken as 25, since it has been shown that for many ensemble problems, the biggest profit in
accuracy is already made with this number of individual classifiers [18]. A standard threelayered backpropagation
network with the tangent sigmoid transfer function is considered. The weights and biases of the neural networks are
initialized randomly and the number of neurons in the hidden node is determined heuristically as inputs
small value of the learning rate (0.15) and a large value of the momentum rate (0.8) are chosen to avoid local
minima. The number of training epochs was 500.
outputs
+
. A
Color texture image
(512x512)
Sub image
(128x128)
Sub image
(128x128)
Randomly selected sub image
(128x128)
RBG
LLLH
HLHH
LL LH
HLHH
LLLH
HLHH
Feature Extraction
Wavelet Entropy
Wavelet Energy
Neural Network Ensembles
Decision
Figure 3. The proposed color texture classification scheme
Page 8
Abdulkadir Sengur
The Arabian Journal for Science and Engineering, Volume 34, Number 2B October 2009
498
4. EXPERIMENTS AND DISCUSSIONS
16 real world RGB color images of size 512 x 512 from different natural scenes are used in the experimental
studies. Figure 4 shows the color texture images. A data base of 1920 color image regions of 16 texture classes of
size 128 x128 was constructed randomly by subdividing each color texture image. 320 of the color texture sub
images were used for training the NNensembles. This means that 20 sample sub images for each texture class are
used for training. 100 sub images for each texture class were constructed for testing. This also means that 1600 sub
images in total were used for testing. For comparison purposes, four different randomly chosen wavelet filters are
used. These filters are Bior 4.4, coif4, sym4, and db7. The following feature vectors were constructed:
(1) Intensity (gray scale) images were obtained from the RGB form, thereby discarding the color information.
The features are extracted according to the step 2 of Section 3. Thus, 6 features are obtained for gray scale
textures (1 gray scale image x 3 wavelet detail images x 2 features (norm entropy and l2 norm energy)).
(2) Each R, G, and B component was wavelet transformed using onelevel decomposition and the proposed
feature extraction scheme was employed. This process constructed a feature vector of size 18 (3 color
channels x 3 wavelet detail images x 2 feature values (norm entropy and l2 norm energy)).
(3) For performance comparison of the proposed method with the energy correlation signatures, which was
explained in Section 2.4. Each R, G, and B color component was wavelet transformed using onelevel
decomposition, and wavelet energy correlation signatures were calculated. This process constructed 18
features.
The experiments are conducted based on the methodology which is illustrated in Figure 3 and the experimental
results are presented at Table 1. The correct classification rates are indicated for all color texture types and the
related wavelet filters. The p parameter which is used for norm entropy function is 1,5. This value is chosen after
many trials. Table 1 indicates the experimental results for gray scale texture images. Flowers1, Clouds, and Misc
texture images are correctly classified for all wavelet filter types. The correct classification rate is 100 % for these
texture images. This high correct classification rate is obtained for these color texture types because of their
homogeneity. On the other hand, the high correct classification rates are not obtained for the rest of the gray scale
texture images. Another important property which can be extracted from the results is that the Bior4.4 wavelet type
is produced much more accurately than the other wavelet filter types. The correct classification rates for each
wavelet filter type of the gray scale texture images varied between 85.4% and 93.2%.
One observation from experimental results is that features and the subsequent classification performance are
significantly improved when color information is added to the texture property. For example, for Bior4.4 wavelet
filter type, the correct classification rate is 93.2% for feature vector – 1 (gray level texture features). This correct
classification rate is increased to 98.9% when the color information is used (feature vector – 2). The correct
classification rates for color texture images are also given in Table 1. The correct classification rates are nearly 100
% for all wavelet filter types. 100 % correct classification rate is obtained for Flowers1, Clouds, Misc, Fabric2,
Fabric3, and Water texture images. Moreover, for Bior4.4, 13 color texture images are recognized with a success rate
of 100%. On the other hand, similar high correct classification rates are obtained with the other wavelet filters type.
Our goal is investigating the color texture classification improvement so we also compare our proposal with the
wavelet energy correlation signatures. As can be seen from Table 1, for the feature vector – 3, the overall correct
classification rate is higher then the feature vector – 1. Only two color texture images (Clouds and Misc) are
classified correctly as 100 % success rate. The correlation signatures have more successful results when Bior4.4 and
coif 4 wavelet types are used.
The proposed color texture classification scheme can be used for color texture segmentation, as we know that
color texture segmentation is more important for computer vision applications and content based image retrieval. For
instance, for segmentation of the color texture images, the wavelet entropy and wavelet energy values are computed
for a small window centered on each pixel of the image, resulting in one dimensional feature vector per pixel. Each
pixel is then assigned to a particular image region by neural network (supervised) or clustering techniques
(unsupervised).
Page 9
Abdulkadir Sengur
October 2009 The Arabian Journal for Science and Engineering, Volume 34, Number 2B
499
Figure 4. Color texture images from left to right and top to bottom: Grass, Flowers1, Flowers2, Bark1, Clouds, Fabric7, Leaves,
Metal, Misc, Tile, Bark2, Fabric2, Fabric3, Food1, Water, and Food2
Page 10
Abdulkadir Sengur
The Arabian Journal for Science and Engineering, Volume 34, Number 2B October 2009
500
Table 1. Results of Texture Classification Using NNEnsembles With Different Wavelet Filters
(for 1920 Image Regions)
Correct classification (%)
Wavelet Filter Types
Bior 4.4
Grass 90
Flowers1 100
Flowers2 87
Bark1 70
Clouds 100
Fabric7 100
Leaves 95
Metal 95
Misc 100
Tile 85
Bark2 84
Fabric2 97
Fabric3 100
Food1 100
Water 100
Food2 88
Grass 100
Flowers1 100
Flowers2 95
Bark1 100
Clouds 100
Fabric7 100
Leaves 100
Metal 94
Misc 100
Tile 100
Bark2 100
Fabric2 100
Fabric3 100
Food1 100
Water 100
Food2 93
Grass 100
Flowers1 82
Flowers2 96
Bark1 70
Clouds 100
Fabric7 91
Leaves 77
Metal 100
Misc 100
Tile 100
Bark2 100
Fabric2 100
Fabric3 100
Food1 100
Water 100
Food2 98
Overall results
Feature Vector – 1 93.2
Feature Vector – 2 98.9
Feature Vector – 3 94.6
Images
Coif 4
82
100
70
70
100
97
85
97
100
99
60
62
94
91
100
80
100
100
100
99
100
100
99
99
100
100
100
100
100
100
100
97
94
71
78
93
100
70
78
83
100
100
100
100
76
97
99
100
Sym 4
88
100
74
85
100
85
98
91
100
98
87
100
92
81
98
98
100
100
100
90
100
91
100
100
100
96
97
100
100
99
100
91
98
78
78
100
100
71
72
71
100
91
92
90
83
99
100
87
Db7
90
100
70
53
100
90
71
89
100
100
91
100
93
75
84
61
98
100
100
88
100
100
97
100
100
100
98
100
100
100
100
97
94
78
84
86
100
72
72
72
100
91
83
100
77
91
98
79
Feature Vector – 1
Feature Vector – 2
Feature Vector – 3
86.7
99.6
89.9
92.9
97.8
88.1
85.4
98.6
86.1
Total
Page 11
Abdulkadir Sengur
October 2009 The Arabian Journal for Science and Engineering, Volume 34, Number 2B
501
5. CONCLUSIONS
In this paper, we have discussed the effect of the color and wavelet domain features on the texture classification
problem. The main aim of the study is combining the color and texture information to improve the classification of
the texture images. We proposed a system which uses the wavelet domain entropy and energy quantities of the red,
green, and blue component of the RGB texture images. Among the three wavelet based methods that we examined,
our proposed system for color texture provides the best classification result.
Experimental studies and subsequent results using a set of real world colored texture images show the usefulness
of the wavelet entropy and energy for color texture analysis. The results show that color is an important component
for improving the classification results for the texture analysis problem.
In this study, several important parameters such as wavelet decomposition level, wavelet filter type, and norm
entropy parameter value are constant. Selecting the best decomposition level is an important issue. Furthermore,
selecting the best wavelet filter type and the best p parameter value for norm entropy will be studied in the future.
Another important point is the chosen color space. Several color spaces, such as KL color space, I1I2I3 color space,
and UVW color space, will be added in future works.
REFERENCES
[1] J. Sklansky, “Image Segmentation and Feature Extraction”, IEEE Trans. System Man Cybernat., 8(1978),
pp. 237–247.
[2] G. Van de Wouwer, P. Scheunders, S. Livens, and D. Van Dyck, “Wavelet Correlation Signatures for Color Texture
Characterization”, Pattern Recognition, 32(3)(1999), pp. 443–451.
[3] R. M. Haralick, K. Shanmugan, and I, Dinstein, “Texture for Image Classification”, IEEE Trans. Systems Man
Cybernet, 3(3)(1973) pp. 610–621.
[4] T. R. Reed and J. M. H. du Buf, “A Review of Recent Texture Segmentation and Feature Extraction Techniques”,
CVGIP: Image Understanding, 57(3)(1993), pp. 359–372.
[5] A. Sengur, I. Turkoglu, and M. C. Ince, “Wavelet Packet Neural Networks for Texture Classification”, Expert
Systems With Applications, 32(2)(2007), pp. 527–533.
[6] A. Sengur, “Wavelet Transform and Adaptive NeuroFuzzy Inference System for Color Texture Classification”,
Expert Systems With Applications, 34(3)(2008), pp. 2120–2128
[7] T. Caelli and D. Reye, “On the Classification of Image Regions by Color Texture and Shape”, Pattern Recognition,
26(4)(1993), pp. 461–470.
[8] S. Arivazhagan, L. Ganesan, and V. Angayarkanni, “Color Texture Classification Using Wavelet Transform”,
Proceedings of the Sixth International Conference on Computational Intelligence and Multimedia Applications
(ICCIMA’05).
[9] S. A. Karkanis, D. K. Iakovidis, D. E. Maroulis, D. A. Karras, and M. Tzivras, “ComputerAided Tumor Detection in
Endoscopic Video Using Color Wavelet Features”, Information Tech. in Biomedicine, IEEE Transactions,
7(3)(2003), pp. 141–152.
[10] M. S. Crouse, R. D. Nowak, and R. G. Baraniuk, “Wavelet Based Statistical Signal processing Using Hidden Markov
Model”, IEEE Trans. Signal Process, 46(4)(1998), pp. 886–902.
[11] X. Qing, Y. Jie, and D. Siyi, “Color Texture Analysis Using the WaveletBased Hidden Markov Model”, Patt. Rec.
Lett., 26(11)(2005), pp. 1710–1719.
[12] I. Daubechies, “Orthogonal Bases of Compactly Supported Wavelets”, Comm. Pure. Appl. Math, 41(1988),
pp. 909–996.
[13] C. M. Bishop, Neural Networks for Pattern Recognition. Oxford: Clarendon Press, 1996.
[14] D. Hanbay, I. Turkoglu, and Y. ve Demir, “An Expert System Based on Wavelet Decomposition and Neural
Network for Modeling Chua's Circuit”, Expert Systems With Applications, 34(4) (2008), pp. 2278–2283.
[15] A. Sengur, I. Türkoğlu, and M. C. ve İnce, “Wavelet Oscillator Neural Networks for Texture Segmentation”, Neural
Network World, 4(2008), pp. 275–289.
[16] J. J. Rodríguez and J. Maudes, “Boosting Recombined Weak Classifiers”, Pat. Rec. Let., 29(2008), pp. 1049–1059.
[17] L. Nanni and A. Lumini, “Ensemblator: An Ensemble of Classifiers for Reliable Classification of Biological Data”,
Pat. Rec. Let., 28(2007), pp. 622–630.
[18] L. Breiman, “Bagging Predictors”, Machine Learning, 24(1996), pp. 123–140.
Page 12
Abdulkadir Sengur
The Arabian Journal for Science and Engineering, Volume 34, Number 2B October 2009
502
[19] F. Phan and E. MicheliTzanakou, Supervised and Unsupervised Pattern Recognition: Feature Extraction and
Computational Intelligence. CRC Press, 2000.
[20] R. R. Coifman and M. V. Wickerhauser, “EntropyBased Algorithms for Best Basis Selection”, IEEE Transactions
on Information Theory, 38(2)(1992), pp. 713–718.
[21] Internet: University of Oulu texture database, http://www.outex.oulu.fi/outex.php, 2005.