TEXTURE CLASSIFICATION BY USING ADVANCED LOCAL BINARY PATTERNS AND

SPATIAL DISTRIBUTION OF DOMINANT PATTERNS

Shu Liao and Albert C. S. Chung

Lo Kwee-Seong Medical Image Analysis Laboratory

Department of Computer Science and Engineering,
The Hong Kong University of Science and Technology, Clear Water Bay, Hong Kong.
{liaoshu,achung}@cse.ust.hk

ABSTRACT invariant features by computing the covariance. Chetverikov

used anisotropy features [2] to classify rotated texture im-
In this paper, we propose a new feature extraction method,
ages. Also, several rotation invariant feature extraction meth-
which is robust against rotation and histogram equalization
ods have been proposed by modifying some existing success-
for texture classification. To this end, we introduce the con-
ful rotation sensitive methods. For example, Kashyap and
cept of Advanced Local Binary Patterns (ALBP), which re-
Khotanzad et al. [3] developed an isotropy circular Gaus-
flects the local dominant structural characteristics of different
sian MRF (ICGMRF) method [3] based on the GMRF method
kinds of textures. In addition, to extract the global spatial dis-
proposed by Chellappa [4]. Deng and Clausi [5] extended the
tribution feature of the ALBP patterns, we incooperate ALBP
ICGMRF approach by using similar circular neighborhoods
with the Aura Matrix measure as the second layer to analyze
so that it is more robust to non-isotropy textures, which is
texture images. The proposed method has three novel contri-
the anisotropic circular Gaussian MRF model (ACGMRF).
butions. (a) The proposed ALBP approach captures the most
Porter and Canagarajah [6] extracted rotation invariant fea-
essential local structure characteristics of texture images (i.e.
tures based on the traditional wavelet transform [7]. They
edges, corners); (b) the proposed method extracts global in-
removed the HH wavelet channels and combined the LH and
formation by using Aura Matrix measure based on the spatial
HL wavelet channels to obtain rotation invariant features. Also,
distribution information of the dominant patterns produced
Ojala et al. [8] proposed the uniformed local binary patterns
by ALBP; and (c) the proposed method is robust to rotation
(LBP) approach to extracting rotation and histogram equal-
and histogram equalization. The proposed approach has been
ization invariant features, which was extended by Huang, Li
compared with other widely used texture classification tech-
and Wang by computing the derivative-based local binary pat-
niques and evaluated by applying classification tests to ran-
terns and applied it to the application of face alignment [9].
domly rotated and histogram equalized images in two differ-
ent texture databases: Brodatz and CUReT. The experimental The approach of the conventional LBP is simple and effi-
results show that the classification accuracy of the proposed cient. However, as the conventional LBP just considers the
method exceeds the ones obtained by other image features. uniform patterns in the images. It discards important pat-
tern information for images whose dominant patterns (e.g. the
Index Terms— Texture Classification, Advanced Local
specific patterns with the largest proportion among all the pat-
Binary Patterns, Spatial Distribution
terns) are not uniform patterns. Also, the features of the con-
ventional LBP are the histograms of the uniform patterns in
1. INTRODUCTION a texture image. As such, the spatial distribution information
Analysis of textures plays an important role in many appli- of the patterns (e.g. the locations of the patterns in the image)
cations in computer vision, for example, image retrieval, face is lost, which will be shown that it is an important feature to
image analysis, and motion analysis. A very challenging prob- classify the textures in this paper.
lem in texture classification is to extract rotation and histogram Therefore, we are motivated to propose a new rotation and
equalization invariant features. This problem is also of par- histogram equalization invariant texture classification method
ticular interest because the applications of rotation and his- by extending the conventional LBP approach to reflect the
togram equalization sensitive feature extraction methods are dominant pattern information contained in the texture images
strictly limited. and capturing the spatial distribution information of the domi-
In the past few decades, some researchers have considered nant patterns. The main contributions of this paper are twofold.
applying various kinds of methods to extract rotation invari- First, based on the framework of the conventional LBP, we
ant texture features. Madiraju et al. [1] extracted rotation extend it by observing the frequency histogram of all the la-

1­4244­0728­1/07/$20.00 ©2007 IEEE I ­ 1221 ICASSP 2007

bels of the patterns and then use the histogram distribution of pattern information in the texture images by just considering
the dominant patterns as features. Therefore, the features of the ”uniform local binary patterns”. Table 1 lists the propor-
the improved LBP method can describe the dominant pattern tions of the ”uniform local binary patterns” with different ra-
structures in a texture image more reliably and effectively. dius R in some texture images in the CUReT database.
Second, we find out the locations of all the dominant pat-
Textures P=8,R=1 P=16,R=2 P=24,R=3
terns in a texture image and then construct a binary image for Loofa0002 52.73 41.34 31.73
each dominant pattern by marking the locations where such Loofa0008 57.11 45.73 38.10
Brown Bread0017 53.18 39.95 24.82
dominant pattern occurs as ”1”, and ”0” otherwise. Therefore, Brown Bread0022 47.69 32.42 20.76
each binary image contains the spatial distribution informa- Concrete 0009 52.83 44.04 35.18
Concrete 0016 47.91 36.04 26.17
tion of its corresponding dominant pattern. Then, we use the Crumpled Paper0015 50.05 46.18 33.06
Gray Level Aura Matrix [10] to extract the spatial information Crumpled Paper0016 57.85 48.82 36.11
features in each binary image. It is found that significant im-
provement can be made by the modified LBP method. Also, Table 1. Proportions (%) of ”uniform LBP” for some samples in
the CUReT database. It shows that, for some kinds of textures, the
by combining the spatial distribution information of the dom-
dominant patterns are not mainly the ”uniform LBP”.
inant patterns, further improvement can be achieved. In the
experiments, it is observed that high classification accuracy
can be achieved even in the tough situation by applying rota- In order to fully describe the dominant patterns contained
tion and histogram equalization to the challenging databases in the texture images, we extend the conventional LBP to
of texture images. We utilize the support vector machine [11] ALBP. Each pattern in the image is assigned a unique label
as the classifier in this paper. The grid search is performed to by the following equation [8]
find out the best values of the parameters in which the highest
classification accuracy can be achieved for each feature. p−1

We evaluate the performance of the proposed approach by LBP (m, R) = u(ti − tc )2i , (1)
performing texture classification experiments in two different i=0
databases: Brodatz [12] and CUReT textures [13]. Excellent where tc is the intensity of the center pixel, ti is the intensity
classification accuracy can be achieved by using the proposed of the neighbor i, and u(x) is the step function. It is obvious
method. This implies that our method can produce robust ro- that the LBP defined in Equation 1 is not rotation-invariant as
tation and histogram equalization invariant features for dis- the intensity value of ti changes when the circle is rotated by
criminating a large number of textures. The features of our a specific angle. Two patterns should be treated as the same
proposed method are also computationally efficient as it does type if one can be obtained from the other through rotating by
not need to perform any image filtering. a certain angle. Therefore, the ALBP pattern group is defined
as,

2. ADVANCED LOCAL BINARY PATTERNS ALBP s (m, R) = min(Cir(LBP (m, R), n)), (2)
where n = 0, 1, ..., p − 1, Cir(x, n) performs a circular anti-
In the conventional LBP approach [8], the image pixels are
clockwise bitwise shift on the p-bit number by n times.
first labelled by thresholding the difference between the cen-
The histogram of the ALBP group is computed from the
ter pixel and its neighbors using the step function u(x) (i.e.
image and sorted in a descending order. Therefore, the first
u(x) = 1 when x≥0 and u(x) = 0 otherwise). The con-
several entries of the histogram are guaranteed to be the dom-
catenation of the neighboring labels is then used as a unique
inant patterns from the image. Our experiments show that
descriptor for each pattern.
around 80% of the patterns in the image are sufficient for rep-
The patterns are uniform if the transitions between “0”
resenting the information of the dominant patterns, and the 20
and “1” are less than or equal to two. For example, 01100000
leading entries are enough to cover 80% of the whole ALBP
and 11011111 are uniform patterns. The histogram of the
histogram in the texture image.
uniform patterns in the whole image is used as the feature
vector [8].
Multiresolution analysis can be achieved by choosing dif- 3. SPATIAL DISTRIBUTION INFORMATION OF
ferent values of m and R, where m denotes the number of DOMINANT PATTERNS
neighboring pixels with respect to the center pixel, and R Although the advanced local binary patterns can more reliably
represents the distance from the center pixel to each of the and effectively describe the dominant pattern information and
neighboring pixels. more robust to random rotation, the spatial distribution infor-
However, for some textures whose dominant patterns are mation of the dominant patterns (SIDP) is still lost. More
not ”uniform local binary patterns”, especially for the textures precisely, by using the ALBP alone, we only know what are
with irregular shapes and edges, the conventional local binary the dominant patterns in a texture image. However, we do
patterns cannot effectively and reliably capture the majority not know where are the locations of such dominant patterns.

I ­ 1222
 introduced. Suppose that we represent an image X with a
rectangular lattice of m × n grids S, where m and n denote
the width and height of X. X now can be represented by S
as,

S = {s = (i, j)|0 ≤ i ≤ m − 1, 0 ≤ j ≤ n − 1}. (3)

Also, a neighborhood system N = {Ns , s ∈ S} is defined,
where Ns is the neighborhood at site s. There are two impor-
tant properties in the neighborhood system N: (a) s ∈/ Ns and
(b) s ∈ Nt if an only if t ∈ Ns . Property (a) implies that site
s is not included in its neighborhood. Property (B) implies
that the neighborhood is symmetric. In this paper, we use a
Fig. 1. Spatial distribution maps of dominant pattern labelled 24 in symmetric 11×11 window with the target pixel at the center
C32, C34, C14 and C17. as the neighborhood system.
Aura: [10] Given two subsets A,B ⊆ S, the aura of
A with respect to B for neighborhood system N , denoted
ϑB (A,N ), is given by,

ϑB (A, N ) = ∪s∈A (B ∩ Ns ). (4)

Aura M easure: [10] With the same notations as in Eq.
Fig. 2. The four textures obtained from the CUReT database. From 4, the aura measure of A with respect to B, denoted by
left to right: C32, C34, C14, C17. m(A, B), is given by:

m(A, B) = |Ns ∩ B|, (5)
Such information actually is a powerful feature to describe
s∈A
the texture image. To illustrate the concepts of SIDP, Fig. 1
where for a given subset A ⊆ S, |A| is the total number of
plots the distribution of a dominant pattern in several texture
elements in A.
images obtained from the Curet [13] database.
Gray Level Aura M atrix (GLAM ): [10] Let N be the
The four texture images in Fig. 2 actually are misclas- neighborhood system over S, and {Si , 0≤ i ≤ G − 1} be the
sified by using only the conventional LBP or ALBP accord- gray level sets of an image over S. Then, the gray level aura
ing to our experimental results. The reason is that although matrix of the image over N , denoted A, is given by,
ALBP can describe the types of dominant patterns in the tex-
ture more reliably, the spatial distribution information of such A = A(N ) = [m(Si , Sj )], (6)
patterns is lost. In the above four textures, they all have the where G is the total number of gray levels in the image, Si =
same sets of dominant patterns with very similar proportions, {s ∈ S | xs = i} is the gray level set corresponding to the ith
and the pattern type labelled with 24 is one of the dominant level, and m(Si , Sj ) is the aura measure between Si and Sj
patterns. As we can see, the numbers of such dominant pat- defined in Eq. 5.
terns in each texture are very close to each other. However, the The resulting four elements in the GLAM are strong fea-
distribution properties of this pattern in these four texture im- tures, which can describe the spatial distribution information
ages are very different with each other. Therefore, the SIDP of the dominant patterns and are used as features to comple-
is a very important property to describe the textures, which ment with the ALBP.
will be further illustrated in the Experimental Results section.
To extract features from the each dominant pattern distri-
bution maps, we use the Gray Level Aura Matrix (GLAM) 4. EXPERIMENTAL RESULTS
approach [10] to extract features from the dominant pattern We have evaluated our proposed method on two different databases
distribution maps. Since now the input image is the dominant with large sets of texture images: (1) 24 textures selected
pattern distribution maps, which can be treated as a binary from the Brodatz album [12]; (2) 47 textures selected from
image with only two gray levels (with value ”1” assign to the the CUReT database [13]. In our experiment, random rotation
locations where they contain such pattern, and ”0” otherwise), (each image in the training set and testing set was rotated by a
it is very efficient to calculate the GLAM from this binary im- randomly generated angle between 0 to 360 degrees) and his-
age with only two gray levels. There are only four elements togram equalization were performed on each sample on both
in the resulting GLAM. Then, these four elements are used the training and testing sets in order to test the robustness of
as features to describe the spatial distribution characteristic of method. Half of the images in each class were used as training
the dominant patterns. sets and the remaining images were used as testing sets. Our
Some basic concepts related to the GLAM [10] are now approach has been compared with eight widely used texture

I ­ 1223
classification methods. In our experiment, the support vector 5. CONCLUSION
machine (SVM) was used as the classifier. The kernel for the In this paper, we have proposed a new advanced local binary
SVM was the Gaussian Radial Basis Function (RBF). pattern approach based on the conventional LBP. Also, we
Experiments on Brodatz Database: The classification ac- find that the spatial distribution information of dominant pat-
curacies of different approaches under different environments terns (SIDP) actually is a very powerful feature for describing
are listed in Table 2. According to the experimental results, the characteristics of the texture image as it includes the lo-
the proposed ALBP approach can already outperform the other cation information of the dominant patterns in the texture im-
eight methods under different conditions. Also, by embed- ages. It has been evaluated by comparing with eight widely
ding the spatial distribution information of dominant patterns used approaches with two databases: Brodatz and CUReT.
(SIDP) with the ALBP, the classification performance is bet- It is experimentally shown that our approach has excellent
ter than using ALBP alone. performance in texture classification and is very robust to his-
Experiments on CUReT Database: In the CUReT database togram equalization and random rotation. Computational sim-
[13], it has the largest number of texture classes (47 classes). plicity is another advantage of our proposed method as the
The major characteristic of this database is that the number features can be obtained with only a few calculations and
of texture classes is very large. Therefore, it is very diffi- comparisons without the need of performing any image fil-
cult to classify such large number of textures because it can tering.
have small inter-class distances in the feature space. This
6. REFERENCES
database actually can test how precise can the features of each
approach describe the texture images. The experimental re- [1] S.V.R. Madiraju and C.C Liu, “Rotation invariant texture classification
sults are listed in Table 3. We can see that, by embedding the using covariance,” in Proc. Int’l Conf. Image Processing, 1994, vol. 2,
pp. 655–659.
SIDP features to the ALBP features, our proposed method
[2] D. Chetverikov, “Experiments in the rotation-invariant texture discrim-
gives very good performance. ination using anisotropy features,” in Proc. Sixth Int’l Conf. Pattern
Recognition, 1982, vol. 18, pp. 975–985.
Classification accuracy %
[3] R.L. Kashyap and A. Khotanzad, “A model-based method for rotation
Features Original Histogram Randomly Histogram Equalized & invariant texture classification,” PAMI, vol. 8, no. 7, pp. 472–481, 1986.
Textures Equalized Rotated Random Rotated
Textures Textures Textures [4] R. Chellappa and S. Chatterjee, “Classification of textures using gaus-
DBWP [7] 98.06 87.73 81.32 62.71
RDBWP [6] 91.67 75.00 91.20 76.83 sian markov random fields,” IEEE Trans. ASSP, vol. 33, no. 4, pp.
TGF [14] 98.61 91.67 83.51 61.36 959–963, 1985.
CGF [14] 90.07 60.28 88.64 58.42
GMRF [4] 96.70 84.33 50.63 40.75 [5] D.A Huawu Deng, Clausi, “Gaussian mrf rotation-invariant features
ACGMRF [5] 95.83 86.52 93.75 81.65
MRH [15] 93.57 70.00 87.35 58.31 for image classification,” PAMI, vol. 26, no. 7, pp. 951–955, 2004.
LBP [8] 97.22 96.30 92.75 91.50
ALBP 98.61 98.61 96.78 96.76 [6] R. Porter and N. Canagarajah, “Robust rotation-invariant texture clas-
ALBP with SIDP 99.85 99.54 99.54 99.54
sification: Wavelet, gabor filter and gmrf based schemes,” in IEE Proc.
Table 2. Performance of different features in the Brodatz Database. Re- Conf. Vision, Image, and Signal Processing, 1997, vol. 144, pp. 180–
sults of our methods are listed in the last two rows. For each test (column), the 188.
highest classification accuracy is highlighted in bold. DBWP: Daubechies [7] A. Laine and J. Fan, “Texture classification by wavelet packet signa-
wavelet packet features; RDBWP: Rotational invariant DBWP; TGF: Tradi- tures,” PAMI, vol. 15, no. 11, pp. 1186–1191, 1993.
tional Gabor filters; CGF: Circular Gabor filters; GMRF: Gaussian Markov [8] M. Pietikainen T. Ojala and T. Maenpaa, “Multiresolution gray-scale
random fields; ACGMRF: Anisotropic circular Gaussian MRFs; MRH: Mu- and rotation invariant texture classification with local binary patterns,”
PAMI, vol. 24, no. 7, pp. 971–987, 2002.
tiresolution histograms; LBP: Uniform local binary patterns; ALBP: Ad-
[9] S.Z.Li X. Huang and Y. Wang, “Shape localization based on statisti-
vanced local binary patterns; SIDP: Spatial Information of Dominant Pat-
cal method using extended local binary pattern,” in IEEE Proc. Conf.
terns. Image and Graphics, 2004, pp. 184–187.
[10] Xuejie Qin and Yee-Hong Yang, “Similarity measure and learning with
gray level aura matrices (glam) for texture image retrieval,” in CVPR,
Classification accuracy % 2004, vol. 1, pp. 326–333.
Features Original Histogram Randomly Histogram Equalized &
Textures Equalized Rotated Random Rotated [11] V.N. Vapnik, Ed., Statistical Learning Theory, John Wiley and Sons,
Textures Textures Textures
DBWP [7] 90.43 61.17 80.13 54.46 New York, 1998.
RDBWP [6] 85.11 39.89 83.75 45.81
TGF [14] 76.06 44.15 63.25 33.56 [12] P. Brodatz, “Textures: A photographic album for artists and designers,”
CGF [14] 60.64 27.66 61.53 36.80
GMRF [4] 46.81 45.27 27.04 35.63
in Dover, 1966.
ACGMRF [5] 77.13 71.26 76.59 68.25
MRH [15] 69.15 38.83 63.68 38.21 [13] S. Nayar K. Dana, B. Ginneken and J. Koenderink, “Reflectance and
LBP [8] 72.87 69.15 70.25 62.58 texture of real-world surfaces,” ACM Trans. Graphics, vol. 18, no. 1,
ALBP 82.98 79.79 84.18 76.02
ALBP with SIDP 97.64 96.57 96.57 95.80 pp. 1–34, 1999.
Table 3. Performance of different features of 64×64 image resolution in [14] G.M. Haley and B.S. Manjunath, “Rotation-invariant texture classifica-
tion using a complete space-frequency model,” IEEE Trans. Img. Proc.,
the CUReT Database. Results of our methods are listed in the last two rows.
vol. 8, no. 2, pp. 255–269, 1999.
For each test (column), the highest classification accuracy is highlighted in
[15] D. Michael H. Efstathios and S. Nayar, “Multiresolution histograms
bold.
and their use for recognition,” PAMI, vol. 26, no. 7, pp. 831–847, 2004.

I ­ 1224