See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/228895050

HSV-based Color Texture Image Classification using Wavelet Transform and

Motif Patterns

Article  in  Journal of Computers · January 2010

CITATIONS READS

14 1,836

4 authors, including:

Jun-Dong Chang Hong-Hao Chen

Institute for Information Industry National Taiwan University
15 PUBLICATIONS   81 CITATIONS    11 PUBLICATIONS   136 CITATIONS   

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Jun-Dong Chang on 23 February 2015.

The user has requested enhancement of the downloaded file.

 HSV-based Color Texture Image Classification using Wavelet Transform
and Motif Patterns

Jun-Dong Chang1, Shyr-Shen Yu1, Hong-Hao Chen2, and Chwei-Shyong Tsai2,＊

1 Department of Computer Science and Engineering,
National Chung Hsing University,
Taichung 402, Taiwan, ROC
[email protected], [email protected]
2 Department of Management Information Systems,
National Chung Hsing University,
Taichung 402, Taiwan, ROC
[email protected], [email protected]*

Received 6 December 2009; Revised 6 January 2010; Accepted 9 January 2010

Abstract. In this paper, a novel color texture image classification based on HSV color space, wavelet trans-
form, and motif patterns is introduced. Traditionally, RGB color space is widely used in digital images and
hardware. However, RGB color space is not accurate in human visual perception and statistical analysis.
Therefore, HSV color space is applied to obtain more accurate color statistics for extracting features. Due to
extracting texture features in color texture images, wavelet transform and motif co-occurrence matrix are
used in HSV color space for feature extraction. According to characteristic of wavelet transform, the horizon-
tal, vertical and diagonal distributions are presented in sub-bands of a transformed image. Then, texture fea-
tures of the horizontal, vertical and diagonal sub-bands are extracted by the motif co-occurrence matrix. After
feature extraction, support vector machine (SVM) is applied to learn and classify texture classes by the ex-
tracted features. From experimental results, the proposed method is better and more correct than recent RGB-
based color texture image classification.

Keywords: HSV color space, texture classification, wavelet transform, motif patterns, co-occurrence matrix,
support vector machine

1 Introduction

Nowadays, texture analysis plays an important role in many image areas, such as geosciences and remote sens-
ing, medical imaging, defect detection, document processing and image retrieval. Texture is a surface structure
formed by uniform or non-uniform repeated patterns. The patterns also can be the perceived surface such as
mineral, metal or wood which have tactile properties, or they could be reflectance on a surface such as color. In
texture analysis, there are related issues such as texture classification, texture segmentation, and texture synthe-
sis which are concerned by many researchers.
In last few decades, lots of texture classification techniques were proposed in these years. Firstly, the first and
second orders statistics [1, 2] and co-occurrence matrix [3] were proposed for obtaining texture features. Further,
model-based method such as Markov random field (MRF) [4, 5], Gibbs transform [6] and linear regression [7]
are used to obtain distribution probabilities of textures on random fields. In addition, the local binary pattern
(LBP) operator was also proposed to discriminate texture patterns by thresholding gray values of the neighbor-
ing pixels with binary codes [8]. Recently, multi-resolution methods such as wavelet transform is widely used
and applied in texture analysis. Arivazhagan et al. proposed to use wavelet co-occurrence features and wavelet
statistical features to discriminate texture classes [9]. Selvan et al. used singular value decomposition (SVD) on
wavelet transform to model a probability density function for texture classification [10]. Moreover, neural net-
works and machine learning are also applied to learn and classify texture classes by wavelet and information
theory features [11-13]. However, most of texture classification methods are proposed to gray-level texture
images. Due to the complexity of color space, existing methods cannot perform well classification in color tex-
ture images. Therefore, it is an important and imperative work to study on the color texture classification.
This paper proposes a HSV-based color texture image classification using wavelet transform and motif pat-
terns. The rest of this paper is organized as follows. Section 2 introduces HSV color space. Section 3 presents

＊ Correspondence author
 Journal of Computers Vol. 20, No. 4, January 2010

wavelet transform and motif co-occurrence matrix for feature extraction. After that, support vector machine is
applied to learn and classify texture images by extracted features in Section 4. Section 5 presents experimental
results and performance comparisons. Finally, concluding remarks are given in Section 6.

2 HSV Color Space

Fig. 1. HSV color model. [18]

In general, RGB color space is a general color space widely used in digital image display and optical instruments.
However, RGB color space is not sensitive to human visual perception or statistical analysis. HSV color space is
a non-linear transform from RGB color space that can describe perceptual color relationship more accurately
than RGB color space. In this paper, HSV color space is adopted in following procedures.
In Fig. 1, HSV color space is formed by hue (H), saturation (S) and value (V). Hue denotes the property of
color such as blue, green, red, and so on. Saturation denotes the perceived intensity of a specific color. Value
denotes brightness perception of a specific color. Thus it can be seen that HSV color space is different from
RGB color space in color variations. When a color pixel-value in RGB color space is adjusted, intensities of red
channel, green channel, and blue channel of this color pixel are modified. That means color, intensity, and satu-
ration of a pixel is involved in color variations. It is difficult to observe the color variation in complex color
environment or content. However, HSV color space separates the color into hue, saturation, and value which
means observation of color variation can be individually discriminated. According to above descriptions about
HSV color space, it can obviously observe that HSV color space can describe color detail than RGB color space
in color, intensity and brightness. In order to transform RGB color space to HSV color space, the transformation
is described as follows:
⎧0, if max = min
⎪
⎪(60o × g − b + 360o ) mod 360o , if max = r
⎪ max − min
⎪
h=⎨ o b−r ,
⎪60 × max − min + 120 ,
o
if max = g
⎪
⎪ o r−g (1)
⎪⎩60 × max − min + 240 ,
o
if max = b

⎧0, if max = 0
⎪
s = ⎨ max − min min ,
⎪⎩ max =1− , otherwise
max
v = max,

where r, g and b denote red, green and blue normalized in value [0, 1]. In order to quantize the range of the h
plane for extracting features specifically, a quantized table is given as follows [14]:
⎧0, if h ∈ [0, 20) U [315,360]
⎪1, if h ∈ [20,50)
⎪
⎪2, if h ∈ [50,75)
⎪
h = ⎨3, if h ∈ [75,155) , (2)
⎪4, if h ∈ [155,195)
⎪
⎪5, if h ∈ [195, 275)
⎪
⎩6, if h ∈ [275,315)

then the quantized hue plane is obtained. After quantizing the original hue plane, the values of s plane and v
plane are normalized to s plane and v plane with value [0, 255] for coming feature extractions.
64
 Chang et al: HSV-based Color Texture Image Classification

3 Feature Extraction

3.1 Wavelet Transform

(a) (b)
Fig. 2. Wavelet transform: (a) Level-1 decomposition, (b) A decomposed
image “Baboon” by Level-1 transform

Wavelet transform (WT) is a multi-resolution decomposition which divides a spatial domain image into four
sub-band images in frequency domain. According to recent researches, wavelet transform can observe the spatial
information of an image. Due to the filter ability of wavelet transform, the input image (signal) is filtered by low
pass and high pass filters. When the input image is filtered by low pass and high pass filters, the filtered signals
represent spatial information within sub-band images. Therefore, texture images are transformed to sub-band
images by wavelet transform before texture feature extraction. The wavelet transform of 1-dimension signal f(x)
is defined as follow:

(Wψ f )( a , b ) = ∫ f ( x )ψ
*
( a ,b ) ( x )d x ,
1 x−b (3)
ψ ( a ,b ) = ψ( ),
a a

where a denotes the scaling factor, b is the translation parameter, and ψ * ( x ) is the transforming function, also
called mother wavelet. Due to applying wavelet transform in a 2-D image, the transforming results are calcu-
lated by using a separable product of 1-D filter to the image:

LL = ⎢ H x ∗ ⎢⎣ H y ∗ I ⎥⎦ ⎥ (b )
⎣ ↓ 2,1 ⎦ ↓1, 2

HL = ⎢ G x ∗ ⎢⎣ H y ∗ I ⎥⎦ ⎥ (b )
⎣ ↓ 2,1 ⎦ ↓1, 2
, (4)
⎢
LH = H x ∗ ⎣⎢ G y ∗ I ⎦⎥ ⎥ (b )
⎣ ↓ 2,1 ⎦ ↓1, 2

HH = ⎢ G x ∗ ⎢⎣ G y ∗ I ⎥⎦ ⎥ (b )
⎣ ↓ 2,1 ⎦ ↓1, 2

where I is the input image, Hx and Hy are low pass filters, Gx and Gy are high pass filters, b ∈ R 2 , * denotes the
convolution operator, ↓2 denotes down-sampling operation. The four sub-band images are denoted as LL, LH,
HL, and HH which include wavelet coefficients to present detailed characteristics of an image. According to
sub-band images LH, HL, and HH, the image information including vertical, horizontal and diagonal features
can be obtained from these three sub-band images respectively. If the image would be processed with further
decomposition, the first level LL sub-band image is decomposed by above procedures. For this reason, the nor-
malized s plane and v plane are decomposed by 1-level wavelet transform that denotes s LH1 , s HL1 , s HH1 and
v LH1 , v HL1 , v HH1 . After that, these six sub-band images and h plane are further extracted texture features by the
motif co-occurrence matrix.

3.2 Motif Co-occurrence Matrix

Originally, the co-occurrence matrix is proposed by Haralick et al. to extract texture features [3]. By scanning
each pixel with four orientations (i.e. 0, 45, 90, and 135 degrees) to calculate probabilities of neighboring pixels’
occurrences, a co-occurrence matrix can be obtained. According to the co-occurrence matrix, texture relations
can be observed by probabilities of corresponding pixel-value pairs. Then, all pixel-value pairs’ probabilities of
65
 Journal of Computers Vol. 20, No. 4, January 2010

the co-occurrence matrix are calculated to obtain 14 texture features (i.e. Angular second moment, contrast,
correlation, sum of squares, inverse difference moment, sum average, sum variance, sum entropy, entropy, dif-
ference variance, difference entropy, information measures of correlation, and maximal correlation coefficient)
for texture analysis. However, the co-occurrence matrix only observes the pixel distribution for an image glob-
ally. Texture discrimination would be weak when the texture content is not regular. In 2004, Jhanwar et al. pro-
posed the motif co-occurrence matrix using in content-based image retrieval [15]. By using the motif co-
occurrence matrix (MCM), the distribution of local area of an image could be noticeably observed and discrimi-
nated. Six motifs (shown in Fig. 3) are different scanning paths to determine which motif does each 2 × 2 grid
belongs to. Each grid is scanned from its top-left pixel and then searches its close pixel by minimum pixel-value
difference. Before scanning, six motifs are labeled as 0 to 5 respectively. For example, in Fig. 4, the top-left grid
contains four pixels {202, 53, 78, 55}, and then the scanning path goes through {202, 78, 55, 53} sequentially.
According to this scanning path, the top-left grid is determined as “U motif” and labeled “2” in a motif map. The
rest of grids are scanned as above procedures to obtain a motif map of this image (shown in Fig 4(b)). Therefore,
the local variance of an image can be observed by the motif map. Further, the motif map is scanned by the co-
occurrence matrix technique to generate a motif co-occurrence matrix to obtain texture features.

Fig. 3. Six motifs with 2 × 2 grids

202 53 149 54 255 255 255 124

78 55 84 52 57 190 186 250
129 68 35 128 160 38 36 255
183 29 140 68 54 31 144 182
176 52 47 43 47 53 145 156
145 38 61 45 40 62 140 176
150 186 98 488 220 211 87 167
99 196 189 174 155 159 151 106
(a) (b)
Fig. 4. Example of scan motifs (a) An 8 × 8 image, (b) A motif map of (a)

Then, a motif co-occurrence matrix P(i, j) which i and j = 0, 1 …, 5 is generated by scanning a whole motif
map. According to the motif co-occurrence matrix, the features such as contrast, energy and entropycan be cal-
culated and obtained by follow equations:
N
Contrast : C = ∑ (i −
i, j =0
j ) 2 P (i , j ),

N
Energy : E = ∑ P (i , j )
i, j =0
2
, (5)
N
Entropy : H = − ∑ P (i , j ) log 2 P (i , j ),
i, j =0

where N denotes the number of motifs (or color bins).

66
 Chang et al: HSV-based Color Texture Image Classification

In this paper, the h plane and six sub-band images s LH1 , s HL1 , s HH1 and v LH1 , v HL1 , v HH1 for s plane and v
plane are scanned to generate its own motif co-occurrence matrix respectively. Then, 21 features are obtained by
calculating these matrices by Eq. (5). These 21 features also can be formed to be a feature set S to input to sup-
port vector machine to learn and classify texture classes. The details of texture classification are described in
Section 4.

4 Texture Classification

In this section, support vector machine (SVM) is applied to train the extracted texture features by following
procedures. Support vector machine (SVM), a kind of supervised learning algorithm used for classification and
regression, is proposed by Vapnik and collaborators [16, 17]. Due to the significant learning and classification
ability, support vector machine is generally applied to bioinformatics, pattern recognition and data mining. Sup-
port vector machine maps input vectors to a higher dimensional space where a maximal separating hyperplane is
constructed. Two parallel hyperplanes are constructed on each side of the hyperplane that separates the data. The
separating hyperplane is the hyper-plane that maximizes the distance between the two parallel hyperplanes.
Therefore, the benefit of SVM is that it can perform well without many training samples. Let a sample dataset be
described as:
S = {(x i , yi ) | i = 1,2,..., l}, x i ∈ R n , yi ∈ R, (6)
where xi = (xi1, xi2, …, xin) is the ith input pattern with n tuples; yi ∈{1, −1}, is the ith label for xi. xi belongs to either
of two classes. Therefore, SVM finds a hyperplane H for S also can be represented as:
H : w ⋅ x + b = 0,
t
i (7)
where t is the transposing operation, w = (w1, w2, …, wn) is the weight vector, and b is a bias. In order to get an
optimal value for (w, b) to maximize the margin ρ which is 1/||w||2 between w．xi + b = 1 and w．xi + b = -1,
the optimal value for (w, b) can be obtained by solving the follow condition:
min 1 ww t ,
2 (8)
s.t. yi (wx it + b) ≥ 1, i = 1,2,..., l.
The condition is also called prime problem, we can use the Karush-Kuhn-Tucker (KKT) theorem and La-
Grange multiplier method to solve the dual problem as follow:
max ∑ α i − 1 ∑ ∑ α iα j yi y j K (x it x j ),
l l l

i =12 i =1 j =1
(9)
s.t. ∑ α i yi = 0, where α i ≥ 0, i = 1,2,..., l ,
l

i =1

where α i denotes LaGrange multipliers associated with Eq. (9), K (x it x j ) is the kernel function. Finally, the opti-
mal decision function f (x) can be defined as:
f (x) = sign( ∑ α i yi x i x t + b).
l
(10)
i =1

In this paper, the extracted feature set S is inputted to support vector machine to learn and classify texture
classes. Then, the following section presents and discusses the performance of the proposed method.

5 Experimental Results
In this section, a color texture image database (University of Oulu texture database, 2005, shown in Fig. 5) is
used in the experiments. Sixteen RGB color texture images of 512 × 512 pixels size are used in the experiments
for different classes. Then, 100 sub-images are generated by randomly dividing a texture image with 128 × 128
pixels size for each class. That means 1,600 sub-images are generated for 16 classes which include 320 sub-
images for training and 1,280 sub-images for testing. First, all of RGB color sub-images are transformed to HSV
color sub-images formed by h, s, and v plane. According to the ranges of HSV are h ∈ [0,360] , s ∈ [0,1] , v ∈ [0,1] ,
these three planes are normalized to h ∈ [ 0,6] , s ∈ [0, 255] , v ∈ [ 0, 255] . Then, the normalized s plane and v plane
are decomposed by wavelet transform to obtain sub-band images denoted as s LH1 , s HL1 , s HH1 and v LH1 , v HL1 , v HH1 ,
respectively. After obtaining sub-band images, the h plane and six sub-band images are scanned to generate the
motif co-occurrence matrices to calculate homogenous features for SVM training and classification.

67
 Journal of Computers Vol. 20, No. 4, January 2010

Fig. 5. Sixteen 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

Table 1. Comparisons of classification rates (%) for three methods.

Methods
Image The Pro-
Type WSF-WCF W-ANFIS
posed
[10] [12]
Method
Grass 96% 86% 92%
Flowers1 99% 90% 98%
Flowers2 96% 89% 90%
Bark1 98% 90% 94%
Clouds 100% 97% 100%
Fabric7 97% 92% 91%
Leaves 100% 92% 98%
Metal 93% 90% 90%
Misc 99% 93% 94%
Tile 97% 95% 94%
Bark2 97% 90% 92%
Fabric2 100% 97% 99%
Fabric3 100% 99% 99%
Food1 99% 95% 97%
Water 100% 100% 100%
Food2 95% 90% 89%

In this experiment, two wavelet based texture classification methods are compared to the proposed method,
which are the WSF-WCF method and the W-ANFIS method. The WSF-WCF method which is proposed by
Arivazhagan et al. uses wavelet statistical features and wavelet co-occurrence features for texture classification.
On the other hand, the W-ANFIS method also uses wavelet transform and information theory to obtain features,
and then adaptive neuro-fuzzy inference system is applied to learn and classify texture classes. In related re-
searches of texture classification, the correct classification rate is calculated by the first order of classified tex-
ture images which is belonged to its true class or not. In Table 1, it can obviously observe that the correct classi-
fication rate of the proposed method is better than the other methods. Due to properties of the HSV color system,
color, intensity and brightness of the texture images can be observed particularly. Comparing to the WSF-WCF
method, the proposed method uses less features to achieve high correct classification rate. On the other hand,
although the W-ANFIS uses fewer features but RGB color space is not sufficient for better perceptual vision in
feature extraction. Also, the training process of W-ANFIS takes more time in selecting optimal neurons and
parameters. In brief, the proposed method achieves better correct classification rate with fewer features and low
time-consuming than the WSF-WCF method and the W-ANFIS method.

68
 Chang et al: HSV-based Color Texture Image Classification

6 Conclusions

In this paper, a HSV-based color texture image classification technique using wavelet transform and motif pat-
terns is proposed. According to perceptual ability of HSV color space, RGB color texture images are trans-
formed to HSV color space. Then, normalized HSV planes are scanned by motifs to generate three motif co-
occurrence matrices to obtain texture features. By using support vector machine, texture images are learned and
classified by extracted features. From experimental results, the proposed method is better than the WSF-WCF
method and the W-ANFIS method in correct classification rates. According to the performance of the proposed
method, it can be applied to medical image recognition such as breast ultrasound images or bone age assessment
for texture analysis in future works.

7 Acknowledgement

This work is supported in part by the Ministry of Education, Taiwan, R.O.C. under the ATU plan.

References

[1] J. S. Weszka, C. R. Dyer, A. Rosenfeld, “A Comparative Study of Texture Measures for Terrain Classification,” IEEE
Transaction on Systems, Man, and Cybernetics, Vol. 6, No. 4, pp. 269-286, 1976.
[2] O. D. Faugeras and W. K. Pratt, “Decorrelation Methods of Texture Feature Extraction,” IEEE Transactions on Pattern
Analysis and Machine Intelligence, Vol. 2, No. 4, pp. 323-332, 1980.
[3] R. M. Haralick, K. Shanmugam, I. Dinstein, “Textural Features for Image Classification,” IEEE Transactions on System,
Man and Cybernetics, Vol. 3, No. 6, pp. 610-621, 1973.
[4] G. R. Cross and A. K. Jain, “Markov Random Field Texture Models,” IEEE Transactions on Pattern Analysis and
Machine Intelligence, Vol. 5, No. 1, pp. 25-39, 1983.
[5] R. L. Kashyap and A. Khotanzed, “A Model based Method for Rotation Invariant Texture Classification,” IEEE Trans-
actions on Pattern Analysis and Machine Intelligence, Vol. 8, No. 4, pp. 472-481, 1986.
[6] H. Derin and H. Elliot, “Modeling and Segmentation of Noisy and Textured Images using Gibbs Random Fields,” IEEE
Transactions on Pattern Analysis and Machine Intelligence, Vol. 9, No. 1, pp. 39-59, 1987.
[7] Z. Z. Wang and J. H. Yong, “Texture Analysis and Classification with Linear Regression Model based on Wavelet
Transform,” IEEE Transactions on Image Processing, Vol. 17, No. 8, pp. 1421-1430, 2008.
[8] H. Zhou, R. Wang, C. Wang, “A Novel Extended Local-binary-pattern Operator for Texture Analysis,” Information
Sciences, Vol. 178, No. 22, pp. 4314-4325, 2008.
[9] S. Arivazhagan and L. Ganesan, “Texture Classification using Wavelet Transform,” Pattern Recognition Letters, Vol.
24, No. 9-10, pp. 1513-1521, 2003.
[10] S. Selvan and S. Ramakrishnan, “SVD-based Modeling for Image Texture Classification using Wavelet Transforma-
tion,” IEEE Transactions on Image Processing, Vol. 16, No. 11, pp. 2688-2696, 2007.
[11] A. Sengur, “Wavelet Transform and Adaptive Neuro-fuzzy Inference System for Color Texture Classification,” Expert
Systems with Applications, Vol. 34, No. 3, pp. 2120-2128, 2008.
[12] A. Sengur, I. Turkoglu, M. C. Ince, “Wavelet Packet Neural Networks for Texture Classification,” Expert Systems with
Applications, Vol. 32, No. 2, pp. 527-533, 2007.
[13] I. Turkoglu and E. Avci, “Comparison of Wavelet-SVM and Wavelet-adaptive network based Fuzzy Inference System
for Texture Classification,” Digital Signal Processing, Vol. 18, No. 1, pp. 15-24, 2008.
[14] X. Zhu, J. Zhao, J. Yuan, H. Xu, “A Fuzzy Quantization Approach to Image Retrieval based on Color and Texture,”
Proceedings of the 3rd International Conference on Ubiquitous Information Management and Communication
(ICUIMC2009), Suwon, Korea, pp. 141-149, 2009.
[15] N. Jhanwar, S. Chaudhuri, G. Seetharaman, B. Zavidovique, “Content-based Image Retrieval using Motif Cooccurrence
Matrix,” Image and Vision Computing, Vol. 22, No. 14, pp. 1211-1220, 2004.
[16] V. Vapnik, The Nature of Statistical Learning Theory, Springer-Verlag, New York, 1995.
[17] V. Vapnik, Statistical Learning Theory, John Wiley, New York, 1998.
[18] http://en.wikipedia.org/wiki/HSL_and_HSV

69

View publication stats