Pattern Recognition Letters 24 (2003) 3197–3203

www.elsevier.com/locate/patrec

Texture segmentation using wavelet transform

a,* b
S. Arivazhagan , L. Ganesan
a
Department of Electronics and Communication Engineering, Mepco Schlenk Engineering College, Amathur (P.O.),
Sivakasi 626 005, Tamil Nadu, India
b
Department of Computer Science and Engineering, Government College of Engineering, Tirunelveli 627 007, India
Received 4 February 2003; received in revised form 31 July 2003

Abstract

Texture analysis such as segmentation and classification plays a vital role in computer vision and pattern recognition
and is widely applied to many areas such as industrial automation, bio-medical image processing and remote sensing.
This paper describes a novel technique of feature extraction for characterization and segmentation of texture at multiple
scales based on block by block comparison of wavelet co-occurrence features. The performance of this segmentation
algorithm is superior to traditional single resolution techniques such as texture spectrum, co-occurrences, local linear
transforms, etc. The results of the proposed algorithm are found to be satisfactory.

2003 Elsevier B.V. All rights reserved.

Keywords: Texture; Texture segmentation; Feature extraction; Wavelet co-occurrence features

1. Introduction images. The local image region, statistics or

property that is repeated over the textured region,
Texture plays an important role in many ma- is called a texture element or texel. It must be
chine vision tasks such as surface inspection, scene noted that the texture has both local and global
classification, surface orientation and shape de- meaning, i.e., it is characterized by invariance of
termination. Texture is characterized by the spatial certain local attributes that are distributed over a
distribution of gray levels in a neighborhood. region of an image.
Though texture is widely used and intuitively ob- Texture analysis is one of the most important
vious, it has no precise definition due to its wide techniques used in the analysis and interpretation
variability. According to Sklansky (1978), ‘‘an of images, consisting of repetition or quasi repe-
image region has a constant texture if a set of its tition of some fundamental image elements (e.g.,
local properties in that region is constant, slowly Raghu and Yegnanarayana, 1996, p. 1625). There
changing or approximately periodic’’. This defini- are three primary issues in texture analysis, such as
tion explains many of the textures found in natural texture classification, texture segmentation and
shape recovery from texture. In texture classifica-
tion, the problem is identifying the given texture
*
Corresponding author. region from a given set of texture classes. As op-
E-mail address: [email protected] (S. Arivazhagan). posed to texture classification, in which the class

0167-8655/$ - see front matter
2003 Elsevier B.V. All rights reserved.
doi:10.1016/j.patrec.2003.08.005
3198 S. Arivazhagan, L. Ganesan / Pattern Recognition Letters 24 (2003) 3197–3203

label of a single homogeneous region is determined avoided if one uses the wavelet transform, which
using distinguishing features derived from the provides a precise and unifying frame work for the
region, texture segmentation is concerned with analysis and characterization of a signal at differ-
automatically determining the boundaries between ent scales (e.g., Unser, 1995). Another advantage
various textured regions in an image (e.g., Jain of wavelet transform over Gabor filter is that the
et al., 1995). low pass and high pass filters used in the wavelet
Analysis of texture requires the identification of transform remain the same between two consecu-
proper attributes or features that differentiate the tive scales while the Gabor approach requires fil-
textures in the image for segmentation, classifica- ters of different parameters (e.g., Chang and Jay
tion and recognition. The features are assumed to Kuo, 1993). In other words, Gabor filters require
be uniform within the regions containing the same proper tuning of filter parameters at different
textures. Initially, texture analysis was based on scales. Later, Kaplan (1999) proposed extended
the first order or second order statistics of textures fractal analysis for texture classification and seg-
(e.g., Haralick et al., 1973; Weszka et al., 1976; mentation and Wang and Liu (1999) proposed
Davis et al., 1979; Faugeras and Pratt, 1980; Chen multi-resolution MRF (MRMRF) parameters for
and Pavlidis, 1983). Then, Gaussian Markov ran- texture classification. Wavelet statistical features
dom field (GMRF) and Gibbs random field (WSF) and wavelet co-occurrence features (WCF)
models were proposed to characterize textures were proposed and effectively used for texture
(e.g., Cross and Jain, 1983; Chellappa and Chat- characterization and classification (e.g., Arivazha-
terjee, 1986; Kashyap and Khotanzed, 1986; Derin gan and Ganesan, 2003).
and Elliot, 1987; Cohen et al., 1991; Manjunath The wavelet transform is a multi-resolution
and Chellappa, 1991). Later, local linear trans- technique, which can be implemented as a pyramid
formations are used to compute texture features or tree structure and is similar to sub-band de-
(e.g., Laws, 1980; Unser, 1986). Then, texture composition. In this paper, texture segmentation is
spectrum technique was proposed for texture carried out by comparing co-occurrence matrix
analysis (e.g., He and Wang, 1990). The above features derived from discrete wavelet transformed
traditional statistical approaches to texture ana- overlapping but adjacent sub-blocks of size
lysis, such as co-occurrence matrices, second order 32 · 32, both horizontally and vertically. The re-
statistics, GMRF, local linear transforms and sults are found to be satisfactory.
texture spectrum, are restricted to the analysis of This paper is organized as follows: In Section 2,
spatial interactions over relatively small neigh- the theory of discrete wavelet transforms is briefly
borhoods on a single scale. As a consequence, their reviewed. The texture segmentation system is
performance is best for the analysis of micro- explained in Section 3. In Section 4, texture seg-
textures only (e.g., Unser, 1995, p. 1549). mentation experimental results for various texture
More recently, methods based on multi-resolu- mosaic images are discussed in detail. Finally,
tion or multi-channel analysis, such as Gabor fil- concluding remarks are given in Section 5.
ters and wavelet transform, have received a lot of
attention (e.g., Unser and Eden, 1989; Bovik et al.,
1990; Chang and Jay Kuo, 1993; Unser, 1995; 2. Discrete wavelet transform (DWT)
Teuner et al., 1995; Haley and Manjunath, 1995;
Manjunath and Ma, 1996; Wu and Wei, 1996; Wavelets are functions generated from one
Raghu and Yegnanarayana, 1996; Van de Wou- single function W by dilations and translations.
wer et al., 1999). But, the outputs of Gabor filter The basic idea of the wavelet transform is to rep-
banks are not mutually orthogonal, which may resent any arbitrary function as a superposition of
result in a significant correlation between texture wavelets. Any such superposition decomposes the
features. Finally, these transformations are usually given function into different scale levels where each
not reversible, which limits their applicability for level is further decomposed with a resolution
texture synthesis. Most of these problems can be adapted to that level (e.g., Antonini et al., 1992).
 S. Arivazhagan, L. Ganesan / Pattern Recognition Letters 24 (2003) 3197–3203 3199

gray level variations, they are statistically charac-

terized by the values in the DWT transformed sub
band images or the features derived from these
sub-band images or their combinations. In other
words, the features derived from these approxi-
mation and detail sub-band images uniquely
characterize a texture. The features obtained from
these DWT transformed images are shown here as
useful for texture analysis, namely segmentation,
Fig. 1. Image decomposition: (a) one level, (b) two level. and are discussed in Section 3.

The discrete wavelet transform (DWT) is iden-

tical to a hierarchical sub band system where the 3. Texture segmentation system
sub-bands are logarithmically spaced in frequency
and represent an octave-band decomposition. By The steps involved in texture segmentation is
applying DWT, the image is actually divided i.e., shown in Fig. 2.
decomposed into four sub-bands and critically Here, texture mosaic images of size N
N are
sub-sampled as shown in Fig. 1(a). These four sub- considered. The analysis is carried out by consid-
bands arise from separate applications of vertical ering sub-images (i.e., block) of size 32 · 32. Each
and horizontal filters. The sub-bands labeled LH1, 32 · 32 sub-image, taken from top left corner of
HL1 and HH1 represent the finest scale wavelet the original image, is decomposed using one level
coefficients i.e., detail images while the sub-band DWT and co-occurrence matrices (C) are derived
LL1 corresponds to coarse level coefficients i.e., for sub-image and detail sub-bands (i.e., LH1,
approximation image. To obtain the next coarse HL1 & HH1 sub-bands) of wavelet decomposed
level of wavelet coefficients, the sub-band LL1 sub-image. Then, from these co-occurrence ma-
alone is further decomposed and critically sam- trices (C), significant WCFs, such as contrast,
pled. This results in two level wavelet decomposi- cluster shade and cluster prominence, are com-
tion as shown in Fig. 1(b). Similarly, to obtain puted using formulae given in Eqs. (1)–(3), as
further decomposition, LL2 will be used. This texture features. In our implementation, the con-
process continues until some final scale is reached. trast feature values, calculated over all the blocks,
The values or transformed coefficients in ap- are subjected to linear normalization in the scale of
proximation and detail images (sub-band images) 0–255, while the cluster shade and cluster promi-
are the essential features, which are as useful for nence features, which found to have very large
texture discrimination and segmentation. Since dynamic ranges, are subjected to logarithmic nor-
textures, either micro or macro, have non-uniform malization in the scale of 0–255 for computational

Texture
Mosaic image
Sub image DWT Feature
(32 x 32) Block (Decomposition) Extraction
Successive Block
Feature Differences
Thinning Post Processing Segmentation Band

Segmented
Texture

Fig. 2. Texture segmentation system.

3200 S. Arivazhagan, L. Ganesan / Pattern Recognition Letters 24 (2003) 3197–3203

simplicity, without affecting their original varia- thickness. The thinned result gives the line of de-
tions. marcation among the different textures present
in the image i.e., thinned lines are exactly aligned
X
N
with texture boundaries. The texture segmentation
Contrast ¼ ði  jÞ2 Cði; jÞ; ð1Þ
i;j¼1 algorithm is given as follows.

X
N
3
Segmentation algorithm:
Cluster shade ¼ ði  Mx þ j  My Þ Cði; jÞ; Input: Texture mosaic image of size N
N .
i;j¼1 Output: Texture segmented image.
ð2Þ
Step 1. Read the texture mosaic image.
Cluster prominence Step 2. Obtain 32 · 32 sub-image blocks, starting
XN from the top left corner.
¼ ði  Mx þ j  My Þ4 Cði; jÞ; ð3Þ Step 3. Decompose sub-image blocks using 2-D
i;j¼1 DWT.
Step 4. Derive co-occurrence matrices (C) for
where original image, and detail sub-bands of
X
N X
N DWT decomposed sub-image blocks.
Mx ¼ iCði; jÞ and My ¼ jCði; jÞ: Step 5. Calculate WCFs such as contrast, cluster
i;j¼1 i;j¼1 shade and cluster prominence from co-
occurrence matrices.
Then, texture segmentation is carried out by Step 6. Calculate the difference between the sums
comparing the normalized co-occurrence features of WCFs of adjacent sub-image blocks.
of discrete wavelet transformed adjacent but This results in segmentation band.
overlapping 32 · 32 sub-image blocks, both hori- Step 7. Apply disk filtering and thresholding tech-
zontally and vertically. Each successive block is niques to remove noise like artifacts, if
differ from the previous one in its spatial location any, in the segmentation band.
by one column or one row, depending on whether Step 8. Apply skeletonizing algorithm to get
the successive block is taken in horizontal or ver- thinned or segmented line of one pixel
tical direction, respectively. Here, the sum of the thickness.
above normalized features of one block is com-
pared with the corresponding sum of features de-
rived from the next block. The difference in feature
values is less when successive blocks belong to the 4. Experimental results and discussion
same texture region and it increases in the texture
border region while it is high when the successive The segmentation technique discussed in the
blocks are from two different texture regions. previous section is applied on six different texture
By carrying out the above block by block feature mosaic images of size 256 · 256, stitched from
comparison both in horizontal and vertical direc- texture images of Brodatz (1966) texture album.
tions, a segmentation band is formed across the The stitched texture mosaic images, shown in Fig.
texture boundaries. When the difference in feature 3(a), consist of (i) leather, straw, grass and wood
values within the same texture region is high, noise textures of square shape in clockwise direction; (ii)
like artifacts or spurious spots appear in the seg- wood texture of square shape at the center of
mented image. This spurious spots are removed by leather texture; (iii) leather texture of circular
applying disk filtering and thresholding techniques shape at the center of wood texture; (iv) leather
(i.e., post processing). Then, the post processed and water textures of triangular shape with sand
segmented band is thinned using a skeletonizing texture at the center; (v) water and sand textures of
algorithm to get segmented line of one pixel square shape with weave texture of circular shape
 S. Arivazhagan, L. Ganesan / Pattern Recognition Letters 24 (2003) 3197–3203 3201

Fig. 3. Texture segmentation results. (a) Texture mosaic images. Texture spectrum technique: (b) segmented band, (c) thinned results.
Proposed method: (d) segmented band, (e) thinned results.
3202 S. Arivazhagan, L. Ganesan / Pattern Recognition Letters 24 (2003) 3197–3203

at the center and (vi) an irregular texture mosaic Acknowledgements

image consisting of leather, wood, and grass tex-
tures. The segmented band images are obtained for The authors are grateful to the Management,
the above six images by applying the segmentation Principal and Head of the Department of Elec-
algorithm, given in Section 3, and are shown in tronics and Communication Engineering of our
Fig. 3(d). Then, thresholding and disk filtering college for their constant support and encourage-
techniques are applied to remove spurious noise ment. The authors also extend their gratitude to
spots. Since the image windows of larger width are the anonymous reviewers who have given very
considered for computation of features from the good suggestions for this better presentation of
texture, the localization of the segmentation bands our manuscript.
are poor. As a result, the detected boundary lines
between different textured regions are thickened. A
pattern matching thinning algorithm is applied for
thinning the detected segmented band images. The References
thinned results are shown in Fig. 3(e). In addition,
Antonini, M., Barlaud, M., Mathieu, P., Daubechies, I., 1992.
for the purpose of comparison, the texture seg- Image coding using wavelet transform. IEEE Trans. Image
mentation bands and the thinned results, obtained Process. 1 (2), 205–220.
using texture spectrum technique, proposed by He Arivazhagan, S., Ganesan, L., 2003. Texture classification using
and Wang (1990) are given in Fig. 3(b) and (c), wavelet transform. Pattern Recogn. Lett. 24 (9–10), 1513–
1521.
respectively. From the figures, it is observed that
Bovik, A., Clark, M., Geisler, W.S., 1990. Multichannel texture
the segmentation bands, shown in Fig. 3(b), are analysis using localized spatial filters. IEEE Trans. Pattern
not very clear and hence, the thinned results are Anal. 12, 55–73.
very poor while the thinned lines, obtained for the Brodatz, P., 1966. Textures: A Photographic Album for Artists
segmentation bands, shown in Fig. 3(d) are exactly and Designers. Dover, New York.
aligned with the boundaries of texture mosaic Chang, T., Jay Kuo, C.C., 1993. Texture analysis and classi-
fication with tree-structured wavelet transform. IEEE
images. Trans. Image Process. 2 (4), 429–440.
Chellappa, R., Chatterjee, S., 1986. Classification of textures
using Gaussian Markov random fields. IEEE Trans.
5. Conclusion Acoust., Speech, Signal Process. ASSP-33 (4), 959–963.
Chen, P.C., Pavlidis, T., 1983. Segmentation by texture using
correlation. IEEE Trans. Pattern Anal. PAMI-5, 64–69.
In this paper, the concept of discrete wavelet Cohen, F.S., Fan, Z., Patel, M.A., 1991. Classification of
transform is presented for applying to textured rotation and scaled textured images using Gaussian Markov
images for decomposing them into detail and random field models. IEEE Trans. Pattern Anal. 13 (2),
approximation regions. Co-occurrence features, 192–202.
Cross, G.R., Jain, A.K., 1983. Markov random field texture
computed out of the wavelet decomposed im-
models. IEEE Trans. Pattern Anal. PAMI-5 (1), 25–39.
ages, are used for texture segmentation. The idea Davis, L.S., Johns, S.A., Aggarwal, J.K., 1979. Texture analysis
behind this proposed method is to exhibit the using generalized co-occurrence matrices. IEEE Trans.
usage of co-occurrence features computed from Pattern Anal. PAMI-1, 251–259.
discrete wavelet transformed images (i.e., WCF) Derin, H., Elliot, H., 1987. Modeling and segmentation of noisy
for texture segmentation. The features are ap- and textured images using Gibbs random fields. IEEE
Trans. Pattern Anal. PAMI-9, 39–59.
proximately the same when the windows or sub- Faugeras, O.D., Pratt, W.K., 1980. Decorrelation methods of
images considered are from the same texture and texture feature extraction. IEEE Trans. Pattern Anal.
different if they are from different textures. Vari- PAMI-1, 323–332.
eties of textures, collected from standard album, Haley, G.M., Manjunath, B.S., 1995. Rotation invariant
texture classification using modified Gabor filters. Proc.
are stitched to form target images which are used
IEEE, 262–265.
for experimentation and it is found that the pro- Haralick, R.M., Shanmugam, K.K., Dinstein, I., 1973. Texture
posed method yield better results than the texture features for image classification. IEEE Trans. Syst. Man
spectrum method, a single resolution technique. Cyb. 8 (6), 610–621.
 S. Arivazhagan, L. Ganesan / Pattern Recognition Letters 24 (2003) 3197–3203 3203

He, D.-C., Wang, L., 1990. Texture unit, texture spectrum, and Teuner, A., Pichler, O., Hosficka, B.J., 1995. Unsupervised
texture analysis. IEEE Trans. Geo-Sci. Remote Sens. 28 (1), texture segmentation of images using Tuned matched Gabor
509–513. filters. IEEE Trans. Image Process. 4 (6), 863–870.
Jain, R., Kasturi, R., Schunch, B.G., 1995. Machine Vision. Unser, M., 1986. Local linear transforms for texture measure-
McGraw Hill. pp. 234–240. ments. Signal Process. 11, 61–79.
Kaplan, L.M., 1999. Extended fractal analysis for texture Unser, M., 1995. Texture classification and segmentation using
classification and segmentation. IEEE Trans. Image Pro- wavelet frames. IEEE Trans. Image Process. 4 (11), 1549–
cess. 8 (11), 1572–1585. 1560.
Kashyap, R.L., Khotanzed, A., 1986. A model based method Unser, M., Eden, M., 1989. Multiresolution feature extraction
for rotation invariant texture classification. IEEE Trans. and selection for texture segmentation. IEEE Trans. Pattern
Pattern Anal. PAMI-8 (4), 472–481. Anal. 2, 717–728.
Laws, K.L., 1980. Rapid texture identification. Proc. SPIE 238, Van de Wouwer, G., Schenders, P., Van Dyek, D., 1999.
376–380. Statistical texture characterization from discrete wavelet
Manjunath, B.S., Chellappa, R., 1991. Unsupervised texture representation. IEEE Trans. Image Process. 8 (4), 592–598.
segmentation using Markov random fields. IEEE Trans. Wang, L., Liu, J., 1999. Texture classification using multires-
Pattern Anal. 13, 478–482. olution Markov random field models. Pattern Recogn. Lett.
Manjunath, B.S., Ma, W.Y., 1996. Texture features for 20 (2), 171–182.
browsing and retrieval of image data. IEEE Trans. Pattern Weszka, J.S., Dyer, C.R., Rosenfeld, A., 1976. A comparative
Anal. 18 (8), 837–842. study of texture measures for terrain classification. IEEE
Raghu, P.P., Yegnanarayana, B., 1996. Segmentation of Gabor Trans. Syst. Man Cyb. SMC-6 (4), 269–286.
filtered textures using deterministic relaxation. IEEE Trans. Wu, W.-R., Wei, S.-C., 1996. Rotation and gray scale trans-
Image Process. 5 (12), 1625–1636. form invariant texture classification using spiral resampling,
Sklansky, J., 1978. Image segmentation and feature extraction. sub band decomposition and hidden Markov model. IEEE
IEEE Trans. Syst. Man Cyb. 8, 237–247. Trans. Image Process. 5 (10), 1423–1433.