Improving Texture Recognition using Wavelets
Ranjan Parekh1 and Nalin Sharda2
1
School of Education Technology, Jadavpur University, Calcutta, India
2
School of Computer Science and Mathematics, Victoria University, Melbourne, Australia
Abstract – Texture is an important perceptual property of
images based on which image content can be characterized
and searched for in a Content Based Search and Retrieval
(CBSR) system. This paper investigates techniques for
improving texture recognition accuracy by using a set of
Wavelet Decomposition Matrices (WDM) in conjunction with
Grey Level Co-occurrence Matrices (GLCM). The texture
image is decomposed at 3 levels using a 2-D Haar Wavelet
and a coefficient computed from the decomposition matrices is
combined with features derived from a set of normalized
symmetrical GLCMs computed along four directions, to
provide improved accuracy. The proposed scheme is tested on
a set of 13 textures derived from the Brodatz database and is
seen to provide accuracies of the order of 90%.
Keywords: Texture recognition, Grey Level Co-occurrence
Matrix, Wavelet decomposition, Content Based Storage and
Retrieval, Pattern Recognition.
1 Introduction
Over the last decade, the use of multimedia content has
increased phenomenally throughout the world. Application
areas such as digital photo albums, computer based training
(CBT) packages, games and home entertainment, online
business and corporate presentations, information kiosks,
simulation packages, on-line galleries and museums, medical
imaging, and geographical information systems have led to
the growth in the number of digital media repositories. As the
size of such repositories grow, an effective and fast
mechanism for retrieval of digital media content from these
repositories assumes fundamental importance. A repository of
media elements without an effective search and retrieval
mechanism is comparable to a library without a catalog –
even though the information is present it is practically
unavailable to somebody with a specific set of search criteria.
Texture refers to visual patterns or spatial arrangement
of pixels that regional intensity or color alone cannot
sufficiently describe. It is difficult to obtain a general
mathematical model for various textures because of the large
variation in their properties. A first example of the derivation
of features using operators is the set of texture energy
measures formulated by Laws in [1]. In [2] the authors derive
texture operators from co-occurrence matrices. A simple
operator for fast discrimination between textures and uniform
regions have been proposed in [3]. Another method similar to
Laws is described in [4]. Here a set of simple masks (vertical,
horizontal, diagonal and anti-diagonal) are applied. Authors
like Tamura [5] made an attempt at defining a set of visually
relevant texture features. This includes coarseness, contrast,
directionality, line-likeness, regularity, roughness. Fractal
functions have received a great deal of attention in recent
years. Pentland [6] reports a high degree of correlation
between fractal dimensions and human estimates of
roughness. Because of this correlation and the natural
appearance of fractal generated textures, Pentland has
proposed fractal functions as texture models. In [7] the
authors describe a parallel algorithm for segmentation using
simultaneous auto-regressive (SAR) random field models and
multi-dimensional cluster analysis. In [8] the author proposes
a two state Markov model to detect texture edges
characterized by changes in first order statistics. Gabor filters
have been used in several image analysis applications
including texture classification and segmentation [9,10].
Bovik et al [9] suggest the restriction of the choice of Gabor
filters to those with isometric gaussians (aspect ratio one). In
[11] the authors have used the one sided linear prediction
(OSP) model, popularly known as auto-regressive (AR)
model, to derive texture descriptors in terms of the prediction
coefficients.
A popular texture recognition technique is based on
Grey Level Co-occurrence Matrix (GLCM), proposed by
Haralick [12] and subsequently used in a number of research
works. The present work investigates techniques for
improving the texture recognition accuracy by using a set of
Wavelet Decomposition Matrices (WDM) in conjunction with
GLCMs. The organization of the paper is as follows: section 2
provides an overview of GLCMs and WDMs, together with
the proposed scheme, section 3 provides experimental results
and section 4 the overall conclusion and the scope for future
research.
2 Proposed Approach
2.1 Grey Level Co-occurrence Matrix (GLCM)
A GLCM indicates the probability of grey level i
occurring in the neighbourhood of grey level j at a distance
d along direction θ in an image. These probabilities create
the co-occurrence matrix M (i , j | d , θ ) . The symmetrical
GLCM is formed by taking the transpose of the GLCM and
adding it to the original. The GLCM is normalized by dividing
each element by the sum of all elements.
 As an example, if A be a section of an image with 2.3 GLCM based classification
corresponding data matrix shown then the GLCM computed
along the horizontal (θ = 0°) with distance offset 1 is given by A texture class consists of a set of member images :
G , while G0 represents the normalized symmetrical version. Ti = {t1 , t2 ,..., tn }i . For each member image, four directional
GLCMs are computed
0 0 1 1 2 2 1 0 4 2 1 0
0 0 1 1  0 2 0 0   4 0 0  Ti = {(t G 0 , t G 45 , t G 90 , t G135 )1 , (t G 0 , t G 45 , t G 90 , t G135 )2 , (5)
1 2
A= ,G =  , G0 =
0 2 2 2 0 0 3 1 24 1 0 6 1 G
..., (t 0 , t G
45 ,t G
90 ,t G
)}
135 n i
     
 2 2 3 3  0 0 0 1  0 0 1 2 
For each directional GLCM, four feature values viz. contrast,
homogeneity, mean and variance, are calculated Each feature
is averaged over the four directional GLCMs, for each
member image.

Directional GLCMs might be computed along three other

Ti = [( t G C , t G H , t G M , t GV )1,i ,( t G C , t G H , t G M , t GV )2,i (6)
directions : vertical (θ = 90º), right diagonal (θ = 45º) left
G G G G
diagonal (θ = 135º ). ,...,( t C ,t H ,t M ,t ) ]
V n ,i

4 1 0 0 6 0 2 0 2 1 3 0 The texture class is then mapped with the boundary values

 2 2 0   4 2 0   2 1 0 
1 1 1 0 1 1 (min / max) of its features for its member images
G45 = , G90 = , G135 =
18  0 2 4 1 24  2 2 2 2 18  3 1 0 2
     
 0 0 1 0   0 0 2 0   0 0 2 0 
Ti = ( t G C ,max , t G C ,min , t G H ,max , t G H ,min , (7)
G G G G
t M ,max ,t M ,min ,t V ,max ,t )
V ,min i

t G C ,max = max( t G C ,1 , t G C ,2 ,..., t G C , n ),

where
tC ,min = min( t G C ,1 , t G C ,2 ,..., t G C , n )

2.2 GLCM based features A test image s j is assigned a weight of 1 for Ti for each of
A set of features derived from four directional normalized the following conditions i.e. its average GLCM feature value
symmetrical GLCMs have been considered here for texture lies within the boundary values of the training set
characterization viz. GLCM Contrast ( GC ), GLCM
Homogeneity ( GH ), GLCM Mean ( GM ) and GLCM t G C ,max,i ≥ s G C , j ≥ t G C ,min,i ;
Variance ( GV ), as defined below. If Gi , j represents the t G H ,max,i ≥ s G H , j ≥ t G H ,min,i ; (8)
element (i , j ) of a normalized symmetrical GLCM, and N the t G
M ,max,i ≥s G
M,j ≥t G
M ,min,i ;
number of grey levels, G G G
t V ,max, i ≥s V,j ≥t V ,min, i

N N
GLCM Contrast : GC = ∑∑ G
i =1 j =1
i , j (i − j )2 (1)
A cumulative weight is calculated over all the four features for
all texture classes being satisfied,

N N
Gi , j
GLCM Homogeneity : GH = ∑∑ 1 + (i − j)
i =1 j =1
2
(2) s j → Ta (ωa ), Tb (ωb ), Tc (ωc ),... (9)

N N The test sample is assigned the class with the maximum

GLCM Mean : GM = M i = ∑∑
i =1 j =1
iGi , j (3) weight, if present, otherwise a class ‘ × ’ (not determinable).

s j → Ta , if ωa > ωb , ωc ,... (10)

N N
GLCM Variance : GV = ∑∑ G
i =1 j =1
i , j (i − M i )
2 (4) s j → ×, otherwise
2.4 Wavelet Decomposition Matrix (WDM)  a1,1 a1,2 ... a1, n 
a a 2,2 ... a2, n 
A Wavelet [13] is a mathematical function used to A=
2,1
matrix Ci, j for data matrix , is
analyze a time dependent signal at different resolutions. The  ... 
Discrete Wavelet Transform (DWT) analyses the signal at  
 am ,1 am,2 ... am, n 
different resolutions by decomposing it into an approximation
defined as
coefficient and a set of detail coefficients. The Haar Wavelet,
proposed by Alfred Haar, transforms a 1-D signal into a set of
1 m (15)
averages and differences Ci , j = C j ,i = ∑{(ak ,i − µi )(ak , j − µ j )}
m − 1 k =1
x = ( x1 , x2 ,..., xN ) → (s1 ,..., sN / 2 | d1 ,..., d N / 2 ) (11) a1,i + a2,i + ... + am ,i
where, µi = , m being the number of rows
m
x2k −1 + x2k x − x2 k in A. A set of Correlation Matrices
where, sk = , d k = 2 k −1 , k = 1,..., N / 2
2 2 ℜ H1 , ℜV1 , ℜ D1 , ℜ H 2 , ℜV2 , ℜ D2 , ℜ H 3 , ℜV3 , ℜ D3 are computed
As an example the Haar Wavelet transform for a 4-
element 1-D signal is given by from the Covariance Matrices, where

Ci , j
1 1 0 0   x1   x1 + x2  ℜi , j = (16)

1 0 0 1 1   x2  1  x3 + x4 
  
(12)
Ci ,i .C j , j
W4 . X =  . =
2  1 −1 0 0   x3  2  x1 − x2 
      The Wavelet combined feature coefficient is computed as,
0 0 1 −1  x4   x3 − x4 

tW = {µ ( B), σ ( B ), µ (ℜ H1 ), σ ( H1 ), µ (ℜV1 ), σ (V1 ),

For a 2-D signal matrix A (N×N) the corresponding Haar
Wavelet transform is defined as µ (ℜ D1 ), σ ( D1 ), µ (ℜ H 2 ), σ ( H 2 ), µ (ℜV2 ), σ (V2 ), (17)
µ (ℜ D2 ), σ ( D2 ), µ (ℜ H3 ), σ ( H 3 ), µ (ℜV3 ), σ (V3 ),
B V (13)
. NT = 
WN . AW
D 
µ (ℜ D3 ), σ ( D3 )}
H
x1 + x2 + ... + xn
where µ ( x1 , x2 ,..., xn ) = and
where, B is the blur or approximation coefficient and n
H n ,Vn , Dn are the horizontal, vertical and diagonal detailed 1 n
coefficients at level n . The B matrix of a specific level is
σ ( x1 , x2 ,..., xn ) = ∑ ( xi − µ )2
n − 1 i =1
used as the data matrix for the next level.
2.6 DWT based classification
2.5 WDM based features
A texture class consists of a set of member images :
 a11 a12  Ti = {t1 , t2 ,..., tn }i . For each member image a Wavelet
The data matrix is partitioned into 2×2 cells   and
 a21 a22  combined feature coefficient is computed.
for each cell the approximation and detail coefficients are
computed as shown. Ti = {t W 1 , t W 2 ,..., t W n }i (18)

B(i, j ) = 14 (a11 + a12 + a21 + a22 )

The texture class is mapped to the boundary values of its
H (i, j ) = 14 {(a11 + a12 ) − (a21 + a22 )} (14) feature for its member images
V (i, j ) = 14 {(a11 + a21 ) − (a12 + a22 )}
Ti = {t W max , t W min }i (19)
D (i, j ) = 14 {(a11 + a22 ) − (a12 + a21 )}

The texture image is decomposed using a Haar Wavelet with A test image s j is assigned a weight of 3 for texture class
3-level decomposition producing following coefficients Ti = {t1 , t2 ,..., tn }i if the its Wavelet coefficient satisfies the
B, H1 ,V1 , D1 , H 2 , V2 , D2 , H 3 ,V3 , D3 . A set of covariance
following condition,
matrices CH1 , CV1 , CD1 , CH 2 , CV2 , CD2 , CH 3 , CV3 , CD3 , are
computed from the detail coefficients where the covariance tW max,i ≥ sW j ≥ t W min,i ; (20)
The test image is then mapped to all the corresponding texture Table 1. GLCM feature values for training set
classes, if present, otherwise a class ‘×’ (not determinable). GLCM Contrast
Class M1(0°) M2(60°) M3(120°) M4(200°)
Bark 360.4260 345.9385 359.7232 339.0615
s j → Ta , Tb , Tc ... (21) Brick 171.4014 179.7685 179.1377 175.1867
Bubbles 324.1034 293.0699 303.2612 321.4721
Grass 1039.9392 1095.8114 929.5233 915.6591
Leather 606.2629 688.7820 685.9577 617.8599
3 Experimentations Pigskin 159.1169 155.6181 153.2743 155.5879
Raffia 159.8644 156.6128 163.4152 170.4513
3.1 Experimental data set Sand 206.6703 210.7270 210.4654 199.5545
Straw 518.4739 547.8408 601.9059 468.5335
Water 90.8227 101.7042 106.4616 92.6248
Texture samples, from the texture database of the Signal Weave 237.4043 236.5369 242.4293 240.1944
and Image Processing Institute, Electrical Engineering Wood 139.8697 167.2208 170.6442 154.4942
Department, University of Southern California, available at : Wool 151.1335 167.9083 177.1072 168.6971
http://sipi.usc.edu/database which in turn have been derived GLCM Homogeneity
from the Brodatz texture database [14], are divided into 13 Class M1(0°) M2(60°) M3(120°) M4(200°)
Bark 0.0851 0.0849 0.0857 0.0852
categories : bark (D12), brick wall (D94), plastic bubbles
Brick 0.1308 0.1268 0.1261 0.1229
(D112), grass (D9), pressed calf leather (D24), pigskin (D92),
Bubbles 0.1185 0.1161 0.1145 0.1267
raffia (D84), beach sand (D29), straw (D15), water (D38),
Grass 0.1002 0.1064 0.0723 0.0776
herringbone weave (D16), wood grain (D68), woolen cloth Leather 0.0578 0.0598 0.0576 0.0579
(D19) (Fig. 1) Pigskin 0.1095 0.1102 0.1108 0.1103
Raffia 0.1305 0.1276 0.1261 0.1244
Sand 0.1171 0.1162 0.1158 0.1175
Straw 0.0778 0.0801 0.0726 0.0823
Water 0.1414 0.1341 0.1308 0.1380
Weave 0.0959 0.0951 0.0939 0.0946
Wood 0.1967 0.1750 0.1705 0.1756
Wool 0.1053 0.1018 0.0991 0.1008
GLCM Mean
Class M1(0°) M2(60°) M3(120°) M4(200°)
Bark 115.0996 114.9565 114.6904 116.3093
Brick 135.8444 134.2834 133.7450 130.8294
Bubbles 79.9512 83.1068 80.3026 75.9412
Grass 91.0985 89.9808 97.4626 96.1444
Fig. 1. Texture classes – bark, brick, bubbles, grass, leather, Leather 90.2151 87.9846 86.7322 88.5559
pigskin, raffia, sand, straw, water, weave, wood, wool Pigskin 124.7919 126.4490 127.1529 126.9458
Raffia 146.2553 146.1539 144.3226 142.9006
3.2 Training phase Sand 129.5074 128.9377 128.7284 127.5578
Straw 108.9450 106.0127 109.9567 109.9994
For each category four member images have been used with Water 117.4929 118.4285 118.8861 120.4955
rotated at angles 0°, 60°, 120° and 200° angles, making a total Weave 163.8639 164.0402 162.3184 161.3729
of 52 images in the training data set. The rotated images for Wood 173.9048 172.5351 171.6212 169.8766
the “brick” texture class are shown in Fig. 2. Wool 139.0376 137.4516 131.4930 133.2385
GLCM Variance
Class M1(0°) M2(60°) M3(120°) M4(200°)
Bark 1961.0157 1921.0227 2003.0859 1931.3037
Brick 966.9865 971.2114 938.6290 894.7406
Bubbles 1830.1217 1760.9854 1731.7169 1848.4516
Grass 2498.5069 2568.9421 2203.1034 2252.6238
Fig 2. Images of “brick” texture class rotated at angles 0°, Leather 1530.4422 1686.3230 1688.2279 1552.9396
60°, 120° and 200° Pigskin 570.5229 573.3804 567.6617 583.0772
Raffia 665.4577 690.9592 705.9475 724.4272
Sand 747.9029 754.5747 730.4221 707.7082
For GLCM based classification, four directional GLCMs are
Straw 1959.2034 2059.4452 1934.5117 1873.3372
computed for each member image and for each GLCM four Water 298.0885 306.9169 322.4701 311.3212
features i.e. GLCM Contrast, GLCM Homogeneity, GLCM Mean Weave 691.4132 681.5290 697.8408 696.7245
and GLCM Variance are computed, which are then averaged over Wood 474.4028 528.3040 530.6977 527.7685
the four directional GLCMs (Table 1). Wool 665.0775 683.7700 718.7822 693.4032
Each texture class is now mapped to the boundary values of
the texture features of its member images (Table 2).

Table 2. GLCM boundary values for training set

Features GLCM Contrast GLCM Homogeneity
Textures Min Max Min Max Fig 3. Images of “brick” texture rotated at 30°, 90° and 150°
Bark 339.0615 360.4260 0.0849 0.0857
Brick 171.4014 179.7685 0.1229 0.1308 The test images are numbered T1 to T39 as detailed below :
Bubbles 293.0699 324.1034 0.1145 0.1267 T1-T3: Bark (Ba), T4-T6: Brick (Br), T7-T9: Bubbles (Bu),
Grass 915.6591 1095.811 0.0723 0.1064 T10-T12: Grass (Gr), T13-T15: Leather (Le), T16-T18:
Leather 606.2629 688.7820 0.0576 0.0598 Pigskin (Pi), T19-T21: Raffia (Ra), T22-T24: Sand (Sa), T25-
Pigskin 153.2743 159.1169 0.1095 0.1108 T27: Straw (St), T28-T30: Water (Wa), T31-T33: Weave
Raffia 159.8644 170.4513 0.1244 0.1305
(We), T34-T36: Wood (Wd), T37-T39: Wool (Wl).
Sand 199.5545 210.7270 0.1158 0.1175
Straw 468.5335 601.9059 0.0726 0.0823
Water 90.8227 106.4616 0.1308 0.1414 For GLCM based classification, average feature values are
Weave 236.5369 242.4293 0.0939 0.0959 computed over the directional GLCMs for each test image
Wood 139.8697 170.6442 0.1705 0.1967 (Table 4)
Wool 151.1335 177.1072 0.0991 0.1053
Features GLCM Mean GLCM Variance Table 4. GLCM feature values for testing set
Textures Min Max Min Max Contrast Hom. Mean Variance
Bark 114.6904 116.3093 1921.023 2003.068 T1 346.2047 0.0851 114.7130 1936.401
Brick 130.8294 135.8444 894.7406 971.2114 T2 353.3469 0.0863 115.9787 1913.306
Bubbles 75.9412 83.1068 1731.717 1848.452 T3 357.4701 0.0832 114.6977 1939.910
Grass 89.9808 97.4626 2203.103 2568.942 T4 174.9632 0.1275 134.0448 982.7670
Leather 86.7322 90.2151 1530.442 1688.228 T5 185.6811 0.1260 132.6855 939.4938
Pigskin 124.7919 127.1529 567.6617 583.0772 T6 172.9089 0.1270 133.3914 899.7543
Raffia 142.9006 146.2553 665.4577 724.4272 T7 285.5223 0.1093 84.5913 1650.212
Sand 127.5578 129.5074 707.7082 754.5747 T8 341.6952 0.1426 74.7443 1957.268
Straw 106.0127 109.9994 1873.337 2061.593 T9 296.5630 0.1139 81.9135 1722.102
Water 117.4929 120.4955 298.0885 322.4701 T10 1046.337 0.1005 91.4870 2511.046
Weave 161.3729 164.0402 681.5290 697.8408 T11 907.120 0.0679 98.8358 2159.088
Wood 169.8766 173.9048 474.4028 530.6977 T12 898.158 0.0733 96.8227 2190.633
Wool 131.4930 139.0376 665.0775 718.7822 T13 639.7699 0.0580 88.6850 1600.617
T14 709.0523 0.0589 86.9247 1726.264
For WDM based classification, the WDM coefficient values T15 635.1569 0.0590 87.7095 1613.179
are computed for each member image (Table 3). T16 161.7524 0.1085 124.2569 587.3597
T17 152.4289 0.1109 127.0048 557.6824
T18 160.2897 0.1093 126.7715 590.3734
Table 3. WDM coefficient values for training set
T19 153.0627 0.1286 146.6842 660.2471
WDM Coefficient
T20 158.4697 0.1314 145.2890 686.2209
Class M1(0°) M2(60°) M3(120°) M4(200°)
T21 171.1037 0.1243 143.5682 732.2644
Bark 913.6300 912.5058 910.2786 923.5560
T22 210.0266 0.1161 128.8770 749.4470
Brick 1080.2051 1066.6327 1062.2822 1039.1673
T23 213.5458 0.1156 128.4866 764.0323
Bubbles 632.7226 657.8325 635.1783 600.4989
T24 213.6579 0.1148 126.9291 751.3758
Grass 721.4903 712.4310 772.4053 761.6843
T25 522.2885 0.0802 107.5714 1963.863
Leather 714.3604 696.2757 686.2770 700.9268
T26 593.9861 0.0727 110.3926 2006.357
Pigskin 990.4948 1003.7784 1009.4092 1007.7481
T27 527.8732 0.0757 107.4674 1947.750
Raffia 1162.2421 1161.3351 1146.6183 1135.3279
T28 93.9866 0.1368 117.4375 301.0095
Sand 1028.3194 1023.8270 1022.0031 1012.7492
T29 107.1548 0.1374 118.3296 313.0684
Straw 865.3096 840.9003 872.3494 873.4186
T30 94.6004 0.1381 120.7939 303.5222
Water 932.2509 939.5882 943.1529 956.1594
T31 230.8108 0.0956 164.7187 674.1464
Weave 1303.0498 1304.3172 1290.5478 1283.0278
T32 242.8838 0.0945 163.4901 697.0086
Wood 1383.5352 1372.3416 1365.0401 1351.1059
T33 240.4315 0.0945 162.6476 693.5086
Wool 1104.3967 1091.9336 1044.2499 1058.2917
T34 145.1555 0.1785 172.7989 499.6091
T35 176.4666 0.1890 172.3334 527.2058
3.3 Testing phase T36 156.7772 0.1773 170.2165 525.9463
T37 158.0872 0.1035 138.3223 675.9245
The test data set consist of 3 images from each category but T38 171.8112 0.1012 136.1968 696.5031
now rotated by angles of 30º, 90º and 150º, making a total of T39 167.5495 0.1013 133.3416 705.4039
39 images. Samples for the “brick” class are shown in Fig. 3.
 Table 6. Wavelet coefficient values for testing set
A test image is assigned a weight of 1 for a texture class if its W. Coeff W. Coeff W. Coeff
average GLCM feature value lies within the boundary values T1 910.5541 T14 687.9205 T27 852.8742
of the training set for that class, otherwise denoted by a ‘×’. T2 920.5881 T15 694.4173 T28 931.7645
T3 910.4604 T16 986.2257 T29 938.6798
Table 5. Class estimates based on GLCM features T4 1064.6843 T17 1008.1668 T30 958.4609
C H M V Pr.Class T5 1053.7383 T18 1006.3202 T31 1309.7588
T1 Ba Ba,Gr Ba Ba,St Ba (4) T6 1059.4419 T19 1165.5127 T32 1300.0202
T2 Ba Gr Ba St Ba (2) T7 669.5608 T20 1154.4182 T33 1293.2070
T3 Ba Gr Ba Ba,St Ba (3) T8 591.0787 T21 1140.6390 T34 1374.4556
T4 Br, Wd Br, Ra Br,Wd × Br (3) T9 648.0169 T22 1023.2567 T35 1370.5595
T5 × Br,Bu,Ra Br,Wd Br Br (3) T10 724.3906 T23 1020.1452 T36 1353.8153
T6 Br,Wd Br,Ra Br,Wd Br Br (4) T11 783.3187 T24 1007.6449 T37 1098.8564
T7 × × × Le Le (1) T12 767.1101 T25 854.2814 T38 1082.0697
T8 Ba × × Ba,St Ba (2) T13 701.9473 T26 875.7973 T39 1058.9825
T9 Bu × Bu × Bu (2)
T10 Gr Gr,Wd Gr Gr Gr (4) Table 7. Class estimates based on WDM features
T11 × × × × × Pr. Class Pr. Class Pr. Class
T12 × Gr,St Gr × Gr (2) T1 Ba (3) T14 Le (3) T27 St (3)
T13 Le Le Le Le Le (4) T2 Ba (3) T15 Le (3) T28 Wa (3)
T14 × Le Le X Le (2) T3 Ba (3) T16 Pi (3) T29 Wa (3)
T15 Le Le Le Le Le (4) T4 Br (3) T17 Pi (3) T30 Wa (3)
T16 Ra,Wd,Wl × × × × T5 Br (3) T18 Pi (3) T31 We (3)
T17 Wd,Wl × Pi × × T6 Br (3) T19 Ra (3) T32 We (3)
T18 Ra,Wd,Wl × Pi × × T7 Le (3) T20 Ra (3) T33 We (3)
T19 Wd,Wl Br,Ra × × × T8 Bu (3) T21 Ra (3) T34 Wd (3)
T20 Pi,Wd,Wl Wa Ra Ra,We,Wl × T9 Bu (3) T22 Sa (3) T35 Wd (3)
T21 Wl Br,Bu Ra Sa × T10 Le (3) T23 Sa (3) T36 Wd (3)
T22 Sa Bu,Sa Sa Sa Sa (4) T11 Gr (3) T24 Pi (3) T37 Wl (3)
T23 × Bu Sa × × T12 Gr (3) T25 St (3) T38 Br (3)
T24 × Bu Pi Sa × T13 Le (3) T26 St (3) T39 Br (3)
T25 St Gr,St St Ba,St St (4)
T26 St Gr,St × St St (3) The Pr. Class columns denote the probable class of the test
T27 St Gr,St St Ba,St St (4) sample with the weight shown in parenthesis. The samples
T28 Wa Wa × Wa Wa (3) correctly classified are shown in bold. The classification
T29 × Wa Wa Wa Wa (3) accuracy is 34 out of 39 i.e. 87.17%
T30 Wa Wa × Wa Wa (3)
T31 × Gr,We X Ra,Wl ×
For a combined GLCM + WDM classification, a sample is
T32 × Gr,We We Ra,We,Wl We (3)
assigned a texture class with a higher weightage, if the class
T33 We Gr,We We Ra,We,Wl We (3)
estimate based on the GLCM and WDM classifications are
T34 Wd Wd Wd Wd Wd (4)
different, or assigned a class ‘×’ if weights for different
T35 Br,Wl Wd Wd Wd Wd (3)
T36 Pi,Wd,Wl Wd Wd Wd Wd (4)
classes are equal.
T37 Pi,Wd,Wl Gr,Wl Wl Ra,Wl Wl (4)
T38 Br,Wl Gr,Wl Wl Ra,We,Wl Wl (4)
 GLCM  s j → Ta , if Ta ≠ Tb & ωa > ωb
T39 Ra,Wd,Wl Gr,Wl Br,Wl Ra,Wl Wl (4)  s j → Ta (ωa )  s → T , if T ≠ T & ω > ω (22)
 DWT  j b a b b a
 s j → Tb (ωb )  s → ×, if T ≠ T & ω = ω
The last column denotes the probable class of the test sample   j a b b a
with the maximum weight shown in parenthesis. The samples
correctly classified are shown in bold. The classification
accuracy is 27 out of 39 i.e. 69.23%. The Combined Cl. column (Table 8) denotes the probable
class of the test sample taking into account both the GLCM
For WDM based classification, Wavelet coefficient values are (GLCM Cl.) and WDM (WDM Cl.) classifications and
computed for each test image (Table 6). A test image is assigning the texture class with the higher weight to the test
assigned a weight of 3 for a texture class if its Wavelet sample. The samples correctly classified are shown in bold.
coefficient value lies within the boundary values of the The classification accuracy is 37 out of 39 i.e. 94.87%
training set for that class, otherwise denoted by a ‘×’ (Table
7).
 Table 8. Combined classification 5 References
Test samp. GLCM Cl. WDM Cl. Comb. Cl.
T1 Ba (4) Ba (3) Ba [1] K. I. Laws, “Textured image segmentation”, Ph.D.
T2 Ba (2) Ba (3) Ba dissertation, Univ. Southern California, Los Angeles,
T3 Ba (3) Ba (3) Ba CA, USCIPI Rep. 940, 1980.
T4 Br (3) Br (3) Br [2] R. W. Conners, M M Trivedi, C A Harlow,
T5 Br (3) Br (3) Br “Segmentation of a high resolution urban scene using
T6 Br (4) Br (3) Br texture operators”, Computer Vision, Graphics and
T7 Le (1) Le (3) Le Image Processing, 25, 1984, pp. 273-310.
T8 Ba (2) Bu (3) Bu [3] I. Dinstein, A. C. Fong, L. M. Ni, K. Y. Wong, “Fast
T9 Bu (2) Bu (3) Bu discrimination between homogeneous and textured
T10 Gr (4) Le (3) Gr
regions”. Proc. 7th international conference on Pattern
T11 × Gr (3) Gr
Recognition, Montreal, Canada, 1984, pp. 361-363.
T12 Gr (2) Gr (3) Gr
[4] R Wang, A R Hanson, E M Riseman, “Texture analysis
T13 Le (4) Le (3) Le
based on local standard deviation of intensity”, IEEE
T14 Le (2) Le (3) Le
T15 Le (4) Le (3) Le
Computer Society conference on Computer Vision and
T16 × Pi (3) Pi Pattern Recognition, Florida, 1986, pp. 482-488.
T17 × Pi (3) Pi [5] H. Tamura, S. Mori, T. Yamawaki, “Textural Features
T18 × Pi (3) Pi corresponding to Visual Perceptions”, IEEE Transaction
T19 × Ra (3) Ra on Systems, Man and Cybernetics, 8(6), 1978.
T20 × Ra (3) Ra [6] A. P. Pentland, “Fractal based description of natural
T21 × Ra (3) Ra scenes”, IEEE Trans. Pattern Analysis and Machine
T22 Sa (4) Sa (3) Sa Intelligence, 6(6), 1984, pp. 661-674.
T23 × Sa (3) Sa [7] A Khotanzad, A Bouarfa, “A parallel non-parametric non-
T24 × Pi (3) Pi iterative clustering algorithm with application to image
T25 St (4) St (3) St segmentation”, Proc. 22nd Asilomar Conf. on Signals,
T26 St (3) St (3) St Systems and Computers, IEEE Computer Society,
T27 St (4) St (3) St Pacific Grove, CA, 1988, pp. 305-309.
T28 Wa (3) Wa (3) Wa [8] N. K. Huang, “Markov model for image segmentation”,
T29 Wa (3) Wa (3) Wa Proc. 22nd Allerton Conf. on Communication, Control
T30 Wa (3) Wa (3) Wa and Computing, Montecello, 1984, pp. 775-781.
T31 × We (3) We [9] A. C. Bovik, M. Clark and W. S. Geisler, “Multichannel
T32 We (3) We (3) We Texture Analysis Using Localized Spatial Filters”, IEEE
T33 We (3) We (3) We Transactions on Pattern Analysis and Machine
T34 Wd (4) Wd (3) Wd
Intelligence, 12(1), 1990, pp. 55-73.
T35 Wd (3) Wd (3) Wd
[10] B. S. Manjunath, R. Chellappa, “A Unified approach to
T36 Wd (4) Wd (3) Wd
boundary detection”, IEEE Trans. Neural Networks,
T37 Wl (4) Wl (3) Wl
4(1), 1993, pp. 96-108.
T38 Wl (4) Br (3) Wl
T39 Wl (4) Br (3) Wl
[11] K. Deguchi, I. Morishita, “Texture Characterization and
texture based partitioning using two dimensional linear
estimation”, IEEE Trans. Computers, 27, 1978, pp. 739-
4 Conclusions 749.
This paper outlines a scheme for image texture recognition [12] R M Haralick, “Statistical and structural approaches to
based on a three level decomposition using Haar Wavelets. Texture”, Proc. IEEE, 67, 1979, pp. 786 – 804.
Firstly the scheme has been shown to be robust to rotational [13] A. Graps, “An Introduction to Wavelets”, IEEE
variation of texture images. Secondly, it demonstrates an Computational Science and Engineering, 2(2), pp. 50-
improvement in recognition accuracy over another popular 61.
scheme using Grey Level Co-occurrence Matrix (GLCM). [14] P Brodatz, “Textures : A photographic album for artists
Further, it has been shown that the recognition accuracy can and designers” (Dover Publications, NY, 1966).
be further improved by combining GLCMs with Wavelet http://www.ux.his.no/~tranden/brodatz.html
Decomposition Matrices. However, additional measures are
required to address the problems related to variations in
brightness, contrast and tonal range of the images. One way to
tackle this problem is via histogram normalization; another
methodology providing scope for further research is to
combine color and texture, e.g. wood textures combined with
its various hues.