International Journal of Computer Science Trends and Technology (IJCST) – Volume 12 Issue 5, Sep - Oct 2024

RESEARCH ARTICLE OPEN ACCESS

AGLV: An Effective Texture Classification Method

Based on Local Binary Pattern
Shrikant Bhosle [1], Prakash Khanale [2]
Research Scholar, Department of Computer Science, DSM College, Parbhani
Professor & HOD, Department of Computer Science, DSM College, Parbhani

ABSTRACT
Local binary pattern (LBP) is local texture descriptor widely used in texture classification due to its computational simplicity.
However, the LBP descriptor has some weakness such as very sensitive to image rotation, image noise and illumination. In
addition to this, LBP consider only sign difference and ignores magnitude difference that reduces its discrimination ability. In
order to overcome these weaknesses of original LBP, this research work proposed new texture classification method based on
LBP called Average Gray Level Value (AGLV). To make AGLV method rotation invariant and illumination invariant image is
divided into 3*3 regions and use average gray level value of each region to calculate Sign Difference, Magnitude Difference,
Region Based Gray Level value, Min-Max value and β value features from image. All these features capture more detail
texture information for rotation invariant and illumination invariant texture classification. The AGLV method use KNN and
SVM classifier for texture classification. The performance of proposed AGLV method is tested using Brodatz, Kylberg and
Kth-Tips database. The experiment result shows that, the proposed AGLV method is rotation invariant and illumination
invariant. It achieves higher classification result as compare to original LBP method. The texture classification result achieved
by AGLV method by using kylberg texture database is 98.00%, brodatz database 92.02% and using Kth-Tips database is
42.50%.
Keywords — Texture, Texture Classification, LBP, Average Gray Level Value, Rotation invariant, Illumination invariant.

and popular texture descriptor due to its low computational

I. INTRODUCTION complexity and ability to differentiate the macro structure of
Texture is one of the basic important features that describe image such as edge, line and spots.
gray level changes in image. It play important role in Local binary pattern has some limitations such as highly
classification of texture image. Texture classification can be sensitive to image rotation, image noise and change in image
defined as, assigning unknown image to one of the known illumination. Consequence of these limitations, LBP is not
class. It consists of two main steps. First step is extracting able to extract robust texture features from image. To
texture features from image and the second step is overcome these limitations, researcher developed various
classification. It compare test image with all training images texture descriptor based on LBP called LBP variants. LBP
and give result as class name. Texture classification has wide riu2[10], LBPriu2/VAR[10], DLBP[11], CLBP[12], MB-
range of application such as face recognition [1], image LBP[13], ALBP[14], MRELBP[15], AFLBP[16], CRLBP[17],
retrieval [2], fabric inspection [3], remote sensing [4], medical FBLBP [18] are example of LBP variants.
image analysis [5] and so on. If the poor and weak texture In this paper, we try to solve the difficulties of original LBP
features used for texture classification then best classifier will method using proposed new texture descriptor called Average
fail to achieve good classification result. However, a common gray level value (AGLV) method. The AGLV method use
problem in texture classification is that, variation in image average value of 3*3 regions of image. It is used to calculate
orientation and image illumination. It can affect the result of sign difference and magnitude difference features. In addition
texture classification. Consequently, many researchers to this, Region Based Gray Level value, Min-Max value and
focused on developing powerful and efficient texture β value is calculated from each region of image. All these
classification method that is rotation invariant and features are used as texture features for rotation invariant and
illumination invariant. Gaussian Markov Random Field [6], illumination invariant texture classification. Experiment result
GLCM [7], Wavelet Transform [8], LBP [10] are some shows that the AGLV method is rotation invariant and
example of texture classification method developed by illumination invariant. It achieves higher classification result
researcher. Among these methods we focused on Local Binary then original LBP method.
Pattern (LBP) because of its less computational complexity. The rest of the paper is organized as follows: sections 2
Before introducing LBP, Ojala et. al.[9] study the gray briefly review on original LBP method. It also gives detail
level difference between pair of gray level to generate information of proposed AGLV method. The experiment
histogram and used as texture descriptor of image. In [10] results using standard texture databases are discussed in
Ojala develop local texture descriptor work only on gray scale section 3. Finally, section 4 concludes the whole paper.
images called Local Binary Pattern (LBP). It uses only signed
difference to generate eight bit binary pattern. LBP is efficient II. AVERAGE GRAY LEVEL VALUE

ISSN: 2347-8578 www.ijcstjournal.org Page 16

International Journal of Computer Science Trends and Technology (IJCST) – Volume 12 Issue 5, Sep - Oct 2024

In this section, we briefly review on original LBP method • Texture information captured by original LBP method
and its limitations. To improve the performance of texture is limited due to use of only sign difference and ignores
classification, this research work proposed new texture magnitude difference.
classification method based on LBP called Average Gray
Level Value (AGLV) method. It has an ability to extract B. Average Gray Level Value (AGLV) method
rotation invariant and illumination invariant texture features Motivated from limitations of original LBP method, this
from image. research work proposed new texture classification method
based on LBP called Average gray level value (AGLV)
A. Brief Review on Local Binary Pattern method. It explores different way to extract local texture
Local Binary Pattern introduced by ojala [10] extracts only features from image effectively. The AGLV method divides
local texture features from greyscale image. It use 3×3 the image into 3*3 regions and performs thresholding
neighbourhood structure and extract local texture features operation on image called Sign Difference.
from each pixel of image. Figure 1 shows the concept of Local The AGLV method modifies the configuration of original
Binary Pattern. It performs thresholding operation between LBP method. It not only uses average based Sign Difference
central pixel and its surrounding pixels (neighbouring pixels). but also use average based Magnitude Difference of
Thresholding means comparison between central pixel value neighbouring pixel. It can be defined as:
and its neighbouring pixel value. The result of comparison is
set to binary 1 if the value of neighbouring pixel is greater
than or equal to the value of central pixel. Otherwise, it is set
to binary 0.
The eight bit binary number is generated shown in figure 1.
Read this binary number in clockwise direction. The LBP
code is obtained by converting binary number to decimal
number. Then decimal number is used to label the central
pixel of image region. After this, histogram of LBP label is /9
computed over whole image and used as texture features of
image. The LBP features of image are calculated by using Where AVG is average intensity value of all pixel in 3*3
following equations. region and Pi (i=1, 2, 3, ------, 9) indicate the pixel of image
region. SD represents Sign Difference and MD represents
symmetric difference between pixel intensity value and
average value of image region. All pixels in image are
encoded by equation 2 and 3. In addition to this, the proposed
method also extract Region Based Gray Level (RBGL), Min-
Max value and β value of each 3*3 region of image to
represent more detail texture information of image.
Where, the function perform thresholding operation, Pc The RBGL and Min-Max value represent the gray level
represent gray level value of the central pixel and Pi represent changes between different regions of image. All regions of
gray level value of neighbouring pixel (i=1, 2, ---, 8). Figure 1 given image is also encoded by region based gray level
shows Local Binary Pattern. (RBGL) and Min-Max value. Both the features are extracted
from each region of image. The RBGL and Min-Max value
can be defined as:

Binary Pattern=01010100
Fig. 1 Local Binary Pattern

The original LBP method has some limitations. These are:

• It is sensitive to image noise, image rotation and Where Pi (i=1, 2, 3, -------, 9) is pixels of 3*3 image region.
illumination. The β value features are calculated by using square root of
• It uses 3*3 neighborhood system to extract local texture sign difference and it can be defined as:
information from image. Little change to central pixel
of image region can change binary pattern drastically.

ISSN: 2347-8578 www.ijcstjournal.org Page 17

International Journal of Computer Science Trends and Technology (IJCST) – Volume 12 Issue 5, Sep - Oct 2024

Where, SD is sign difference features of each 3*3 regions Performance of proposed AGLV method is evaluated using
of image. following formula:
The MD, SD, β value features are used for rotation
invariant texture classification and Min-Max value, RBGL
features are used for illumination invariant texture
classification.

The block diagram of texture classification algorithm using D. Texture Classification

AGLV method is shown in figure 2. It contains three This Research work use Kylberg texture database [19] to
important Steps such as, Pre-processing Module, Feature test the performance of proposed AGLV method against
Extraction Module and Classification Module. The pre- normal texture image. The kylberg texture database is divided
processing module removes the noise and enhances the image into two parts training and testing. Figure 3 shows the sample
quality. It also changes the scale of both training and test images of kylberg texture database.
image. Feature extraction module divide the image into 3*3
regions and extract Sign Difference (SD), magnitude
difference (MD), RBGL, Min-Max value, -value features
from each region of image. All these features are combining
together and used to generate the histogram. Finally,
histogram features are used as input to the classification
module. The K-Nearest Neighbours and Naïve Bayes
classifier use these histogram features to generate the result.

Fig. 3 Sample Images of Kylberg Texture Database

The kylberg texture database contains 28 classes of texture

images. Due to large size of image database this research
work randomly select only 11 classes of texture images. All
images in this database are gray scale JPEG format images.
The result of AGLV method using Kylberg texture database is
given in table I.
TABLE I PERFORMANCE OF AGLV METHOD USING KYLBERG DATABASE

kylberg database KNN NB

Classifier Classifier
Sr.No. Image Class Performance Performance
Rate Rate
1 Blanket1 100.00% 100.00%
2 Blanket2 100.00% 20.00%
3 Canvas1 100.00% 100.00%
4 Cealing1 100.00% 100.00%
5 Cealing2 100.00% 100.00%
6 Cushion1 100.00% 100.00%
7 Floor1 100.00% 100.00%
Fig. 2 Block Diagram of Texture Classification 8 Floor2 100.00% 100.00%
9 Grass 80.00% 100.00%
III. EXPERIMENT RESULT
10 Rice1 100.00% 100.00%
To evaluate the performance of proposed AGLV method
Average 98.00% 92.00%
experiment are carried out on three standard texture databases
namely, Kylberg, Brodatz and Kth-Tips. The proposed AGLV
method is developed in MATLAB R2015 and uses KNN and
NB classifier for classification of texture image. Both the
classifier used in this research work use different scale of
images. The experiment is conducted as follows:
C. Performance Rate

ISSN: 2347-8578 www.ijcstjournal.org Page 18

International Journal of Computer Science Trends and Technology (IJCST) – Volume 12 Issue 5, Sep - Oct 2024

TABLE II PERFORMANCE OF AGLV METHOD USING BRODATZ DATABASE

Brodatz Database KNN NB

Classifier Classifier
Sr.No. Image Class Performance Performance
Rate Rate
1 Bark 50.00% 83.33%
2 Brick 100.00% 100.00%
3 Bubbles 66.66% 100.00%
4 Grass 100.00% 100.00%
5 Leather 100.00% 83.33%
6 Pigskin 33.33% 66.66%
Fig. 3 Graphical representation of performance of AGLV method
7 Raffia 50.00% 100.00%
The result of AGLV method using KNN classifier is 8 Sand 00.00% 100.00%
98.00% and it is higher as compare to NB classifier. The
result given in table 4.1 proves that, the AGLV method is 9 Straw 50.00% 66.66%
significantly outperform against texture classification using 10 Water 50.00% 83.33%
normal texture image. 11 Weave 00.00% 100.00%
E. Rotation Invariance 12 Wood 100.00% 100.00%
All Brodatz texture database [20] is widely used for 13 Wool 83.33% 100.00%
rotation invariant texture classification. It has 13 classes of Average 60.23% 91.02%
texture images. Each class contains 7 images captured in
different rotation angle. A classifier is trained using images
captured in 0o angle from each class and testing is performed
using texture image rotated in six different angles. Figure 4
shows the sample images of brodatz texture database. The
result of AGLV method using Brodatz database is given in
table II.

Fig. 5 Graphical representation of AGLV method using Brodatz Database

Fig. 4 Sample images of Brodatz Database
Experiment result prove that, the AGLV method achieves
higher classification result 91.02% using NB classifier and it
is higher as compare to KNN classifier. From table 4.3 it is
clear that, the AGLV method is much more robust as compare
to original LBP method. it achieves higher result in rotation
invariant texture classification using Brodatz texture database.
F. Illumination Invariance
To test the illumination invariance of proposed AGLV
method Kth-Tips texture database[21] is used. This database
Fig. 6 Sample images of Kh-Tips Database
contains 10 classes of texture images and all images are
captured in three different illumination condition (from the
front, from the side at roughly 450 and from the top at roughly
450). In addition to this, all images in this database are
captured under 3 different poses (frontal, rotated left 22.50
and rotated right 22.50). There are total 81 images for each
texture class. Figure 3.6 shows the sample images of Kth-Tips
database.

ISSN: 2347-8578 www.ijcstjournal.org Page 19

International Journal of Computer Science Trends and Technology (IJCST) – Volume 12 Issue 5, Sep - Oct 2024

To test the illumination invariance of proposed AGLV approach for texture classification with high discrimination
method, this research work uses only nine images of one scale ability. It gives impressive result for rotation invariant and
from each class. The training images contain only 10 images illumination invariant texture classification.
(one image from each class) and testing images contain only
80 images (8 images from each class). Table III list the REFERENCES
experiment result of AGLV method using Kth-Tips database. [1] J.Y. Choi, Y.M. Ro, and K. Plataniotis, Color local
TABLE III PERFORMANCE OF AGLV METHOD USING KTH-TIPS DATABASE texture features for color face recognition, IEEE
Transactions on Image Processing 21(3) (2012) 1366–
Kth-Tips Database KNN NB
1380.
Classifier Classifier
Sr.No. Image Class Performance Performance
[2] S. Murala, R. P. Maheshwari, and R. Balasubramanian,
Rate Rate Local tetra patterns: A new feature descriptor for
1 Class 06 75.00% 00.00% content-based image retrieval, IEEE Transactions on
Image Processing 21(5) (2012) 2874–2886.
2 Class 15 100.00% 87.50%
[3] F.S. Cohen, Z. Fan, S. Attali, Automated inspection of
3 Class 20 00.00% 00.00% textile fabrics using textural models, IEEE Transactions
4 Class 21 37.50% 12.50% on Pattern Analysis and Machine Intelligence 13 (8)
(1991) 803–808.
5 Class 42 62.50% 37.50%
[4] H. Anys, D.C. He, Evaluation of textural and
6 Class 44 25.00% 50.00% multipolarization radar features for crop classification,
7 Class 46 50.00% 00.00% IEEE Transactions on Geoscience and Remote Sensing
33 (5) (1995) 1170–1181.
8 Class 48 25.00% 37.50%
[5] [5] Q. Ji, J. Engel, E. Craine, Texture analysis for
9 Class 55 50.00% 37.50% classification of cervix lesions, IEEE Transactions on
10 Class 60 00.00% 25.00% Medical Imaging 19 (11) (2000) 1144–1149.
[6] Mike Pereira , Nicolas Desassis , Efficient simulation of
Average 42.50% 28.70%
Gaussian Markov random fields by Chebyshev
polynomial approximation, published by Elsevier,
Volume 31, June 2019
[7] R.M. Haralik, K. Shanmugam, I. Dinstein, Texture
features for image classification, IEEE Transactions on
Systems, Man, and Cybernetics 3 (6) (1973) 610–621.
[8] T. Chang, C.C.J. Kuo, Texture analysis and
classification with tree-structured wavelet transform,
IEEE Transactions on Image Processing 2 (4) (1993)
429–441.
[9] T. Ojala, M. Pietikainen, A Comparative study of texture
measures with classification based on feature
distributions, published by Elsevier (1995).
[10] T. Ojala, M. Pietikainen, T. maenpaa, Multiresolution
gray-scale and rotation invariant texture classification
Fig. 7 Graphical representation of AGLV method using Kth-Tips Database with local binary pattern, IEEE Transactions on Pattern
Analysis and Machine Intelligence 24 (7) (2002) 971–
The AGLV method using KNN classifier gives 42.50% 987.
results and it is higher as compare to NB classifier. The [11] S. Liao, M. Law, Chung, Dominant local binary pattern
AGLV method easily captured illumination invariant texture for texture classification, IEEE Transactions on Image
information. From the table 4.1 it is clear that, the AGLV Processing, 18(5), (2009) 1107-1118.
method using Kth-Tips database perform better for [12] Z. Guo, L. Zhang, D. Zhang, A completed modeling of
illumination invariant texture classification. local binary pattern operator for texture classification,
IEEE Transactions on Image Processing, 19(6) (2010)
IV. CONCLUSIONS 1657-1663.
This research work address the limitations of LBP method [13] L. Zhang, R. Chu, S. Xiang, S. Liao, S. Li, Face
and proposed new method based on LBP called Average Gray detection based on multi-block LBP representation, in
Level Value method for texture classification. it capture more IAPR/IEEE International Conference on Biometrics
detail local texture information from image. Experiment result (2007) 11-18.
obtained from three standard texture databases clearly shows [14] Z. Guo, L. Zhang, D. Zhang, S. Zhang, Rotation
that, the AGLV method gives an effective and efficient invariant texture classification using adaptive LBP with

ISSN: 2347-8578 www.ijcstjournal.org Page 20

International Journal of Computer Science Trends and Technology (IJCST) – Volume 12 Issue 5, Sep - Oct 2024

directional statistical features, in 17th IEEE int. [18] Z. Pan, Z. Li, H. Fan, Feature based local binary pattern
Conference on Image Processing (2010) 285-288. for rotation invariant texture classification, published by
[15] L. Lao, P. Fieguth, Y. Guo, Median Robust Extended Elsevier (2017).
Local Binary Pattern for Texture Classification, IEEE [19] G. Kylberg, The Kylberg Texture Dataset v. 1.0, Centre
Transactions on Image Processing, (2016). for image Analysis, Swedish University of Agricultural
[16] G. Girish, J. Dash, Adaptive Fuzzy Local Binary Pattern Sciences and Uppsala University, External report (Blue
for Texture Classification, in 2nd int. conference on man series) No. 35.
and machine interfacing (2017). [20] P. Brodatz, Textures: A Photographic Album for Artists
[17] Y. Zhao, W. Jia, R. Hu, Completed robust local binary and Designers, Dover Publications, 1966.
pattern for texture classification, published by Elsevier [21] M. Fritz, E. Hayman, B. Caputo, and J.-O. Eklundh, The
(2012). KTH-TIPS database, Available at
www.naa.kth.se/cvap/databases/kth-tips

ISSN: 2347-8578 www.ijcstjournal.org Page 21