Download as pdf or txt
Download as pdf or txt
You are on page 1of 6

RESEARCH PAPERS

EXTENDED RULE BASE LOCAL BINARY PATTERN TECHNIQUE FOR


TEXTURE CLASSIFICATION AND ANALYSIS
By

O. RAMA DEVI * E.V. PRASAD *** L. S. S. REDDY **


* Assistant Professor, PVP Siddhartha Institute of Technology, Vijayawada..
** Pro.VC, K.L University, Vaddeswaram.
*** Director, LBR college of Engineering, Mylavaram.

ABSTRACT
Texture analysis is one of the important and most useful tasks in image processing applications. Several texture models
have been developed over the past few years and Local Binary Patterns (LBP) are one of the simple and efficient
approaches among them. The main disadvantage of “LBP” method is the complex computation of vector generation.
Here an innovative classifier, called Extended Rule Based Local Binary Pattern (ERLBP) is given, which is an efficient model
and its performance has been compared with other widely used texture models to show the computational superiority,
robustness to gray scale variations and improved discriminating capability.
Keywords: Local Binary Patterns, Texture Analysis, Texture Classification, RLBP.

INTRODUCTION parameter estimation is the major problem in stochastic


Image analysis is a process to extract useful information model. Transform methods of texture analysis, such as
from a digital image. Image segmentation, image Fourier [1], Gabor filter [5] represent an image in a space
classification, image correspondence, image domain whose coordinate system has an interpretation
compression are some of the examples of image that is closely related to the characteristics of a texture
analysis. Feature extraction is used to extract some (frequency or size). Methods based on the Fourier
important features like colour, texture and shape from a transform perform poorly in practice, due to its lack of
digital image. Texture analysis is a process of extracting spatial localization. Gabor filters provide a means for
information from the images, which characterizes the better spatial localization; however, their utility is defined in
spatial distribution of the pixel values in the image. practice because, there is normally no single filter
resolution at which one can localize a spatial construction
Literature Review and Related Work
in natural textures Statistical methods analyze the spatial
Texture analysis method can be categorized into a
distribution of pixel values in the digital image based on
number of groups, namely structural, model-based,
the number of pixels defining the local feature
transform methods and statistics. Structural approaches
[6].Statistical methods are the most widely used because
[4], represent texture by well-defined primitives (micro
they can be instantly applied onto any type of grain. Gray-
texture) and a hierarchy of spatial systems (macro texture)
level co-occurrence matrices [4], gray-level differences
of those primitives. Structure based texture is really
[11] and Local Binary Patterns [10] are some of the popular
uncommon in nature and extraction depends on the
statistical texture methods. The Local Binary Pattern
texture primitives and placement patterns. Model based
approach was proved simple yet powerful approach to
texture analysis [3], using fractal and stochastic models
describe local structures [7].
interpret an image texture by use of generative image
1. Local Binary Pattern (LBP)
models and stochastic model. The parameters of the
model are estimated and then applied for image Local Binary Pattern a nonparametric method that
analysis. In exercise, the computational complexity in summarize local structures of images efficiently by

24 i-manager’s Journal on Pattern Recognition, Vol. 1 l


No. 1 l
March - May 2014
ARTICLE
RESEARCH PAPERS

comparing each pixel with its neighbouring pixels [10]. that encodes the sign of GD. The following layers of ELBP
The most important property of LBP is its tolerance with then encode the absolute value of GD as shown in Figure
respective to monotonic illumination changes and its 2(d). Basically, each absolute GD value is first encoded in
computational simplicity. its binary representation, and then all the binary values at
The original LBP operator labels the pixels of an image by a given layer result in an additional LBP representing the
means of decimal numbers, which are called LBP codes sign bits written from bottom to top. For example, in Figure
that encode the local structure S3x3 around each pixel as 2(d), the first layer(L1) is the original LBP code that encodes
shown in Figure 1(a). Every pixel is compared with the the sign of GD, thus yielding a decimal number of 196
center pixel value and negative values are encoded with from its binary form (11000100)2. The absolute values of
0 and the positive with 1 as shown in Figure 1(b). For each GD, i.e., 2, 1, 0, 4, 2, 3, 2, and 1, are representing the
given pixel, a binar y number is obtained by remaining encoded binary numbers: (010)2, (001)2,
concatenating all these binary values in a clockwise (000)2, (100)2,. . ., etc. Using the same weight scheme of
direction starting from top-left neighbour. The LBP on all the binary bits, its ELBP code of the
corresponding decimal value of the generated binary corresponding layer can be generated, e.g., L2 is
number, is used for labelling the given pixel as LBP code, composed of (00001000)2, and its decimal value is 16; L3
as shown in Figure 1(c). is composed of (01110001)2, and its decimal value is
241;and finally, L4 is composed of (10100010)2, and, its
1.1 Recent Developments of Local Binary Pattern
decimal value is 69. As a result, when describing similar
LBP methodology has been developed recently with
local textures, even though the first layer LBP is not
large number of variations for improved performance in
discriminating, the information encoded in the other
different applications. These variations focus on
layers can be utilized to distinguish them. Its downside is
improvement of its discriminating capability. The
that ELBP increases feature dimensionality to a large
encoded local binary pattern (ELBP) [2] is one such
extent and hence computationally more complex.
approach where ELBP operator not only performs a
2. Rule Based Local Binary Patterns
binary comparison, but also encodes their exact Gray-
value Differences (GDs) as shown in Figures 2 (a) and 2(b). Rule Based LBP can be generally described as discrete
The corresponding column vector shown in Figure 2(c). dynamic systems completely defined by a set of rules in a
Specifically, the ELBP feature consists of several LBP codes local neighbourhood. The state of a system is represented
at multiple layers, which encode the GD between the as a regular grid, on which the rules apply to produce a
central pixel and its neighbouring pixels as shown in Figure new state. An interesting property of RLBP is that very
2 (d). The first layer of ELBP is actually the original LBP code simple rules can result in very complex behaviour. Now

Figure 1(a) Figure 1(b)

Binary: 11010011

Decimal: 211

Figure 1©

Figure 1. Basic LBP operator Figure 2. ELBP operator

i-manager’s Journal on Pattern Recognition, Vol. 1 l


No. 1 l
March - May 2014 25
ARTICLE
RESEARCH PAPERS

consider sample window S3X3 and compare each pixel space B3x3
with the significant centre of the sample window. All a) Calculation of the column wise count on sample
neighbouring pixels with values greater than the centre to space B3X3
be replaced with the value 1 otherwise, replace them with
Ci= ∑B (i, n); where n=1, 2,
0 such that, gray images are converted to binary image.
b) Calculation of the row wise count on sample space
1) Calculation of the column wise count on sample space
S3X3
S3X3.
Ri= ∑B (m, i) ; where m=1, 2,3
2) Calculation of the row wise count on sample space S3X3.
c) Calculate the count for left diagonal
3) Calculate the count for left diagonal D1 and right
D1(B 1 1 +B 2 2 +B 3 3 )and right diagonal
diagonal D2 on sample space S3x3.
D2(B13+B22+B31) on sample space S3x3
4) Form the new sample matrix as shown below,
d) Form the new sample matrix as shown below.

5) Compute the LBP operator on new sample window and


Represent matrix as a column vector [R1 D1 C1 C2 C3 D2 R3
replace the centre pixel.
R 2]
6) Repeat step 1 through step5 on entire image to form
e). Convert each value of the column vector into
the new encoded image.
corresponding binary & construct matrix M(the size is
3. Proposed Method (ERLBP) always 8x2)).
Rule based local binary patterns can be generally f) Compute the LBP operator for each of the two columns
described as discrete dynamic systems completely & calculate mean of all 2 LBP codes and replace the
defined by a set of rules in a local neighbourhood. The centre pixel with this value.
state of a system is represented as a regular grid, on which
4) Repeat step 1 through step 3 on the image and form
the rules apply to produce a new state. An interesting
the new unambiguous feature vector of the image.
property of ERLBP is that very simple rules can result in very
4. Experimental Results
complex behaviour. Now consider sample window S3X3
In order to verify the effectiveness of the segmentation
and compare each pixel with the significant centre of the
process of ERLBP, a group of images of different kinds of
sample window. On the image the following rules have
soil images have been tested. The performance
been used to remove the uncertainty of texture
evaluation of ELBP and ERLBP has been described through
classification.
tables of values and graphs. Three texture images of
1. Obtain Encoded image (B) 3x3 from sample gay
sandpebbels, crack and beach of different soil images
image S3x3
are shown in Figures 3,4 and 5, each of size 256X256 were
2. Compute Bi value with Xc & Xi, where Xc is centre of
considered for experiment with 10 samples of each
S3x3 and Xi is the neighbourhood value of S3x3 where Xi≠
image. The resultant values for ELBP and ERLBP for each
Xc.
type of image is shown in Table1 and Table 2. The
Bij= (Sij-Xc) and Bii=Xc Where i, j=1 to 8 classification of ELB for three types of images is shown in
2. Compute mean value of B3x3 and convert sample figure 6 & figure 7. The classification of ERLB for three types
values into binary values based on this mean value as a of images is shown in figure 8 & figure 9. The classification
Threshold. performance of the ERLBP and ELBP can be inferred that
3. Apply RLBP [O.Rama Devi ,13] as below, to count no. Of ELBP texture model cannot differentiate among three
1's in column wise, row wise and diagonal wise on sample categories. ERLBP undoubtedly good texture classifier as it

26 i-manager’s Journal on Pattern Recognition, Vol. 1 l


No. 1 l
March - May 2014
ARTICLE
RESEARCH PAPERS

IMAGE SPERLBP CERLBP BERLBP


is able to clearly distinguish among them.
I1 0 . 221 0 . 231 0 . 235

I2 0 . 226 0 . 242 0 . 247

I3 0 . 228 0 . 244 0 . 236

I4 0 . 231 0 . 244 0 . 244

I5 0 . 236 0 . 245 0 . 252

I6 0 . 243 0 . 246 0 . 262

I7 0 . 245 0 . 25 0 . 263

I8 0 . 245 0 . 25 0 . 266
Figure 3. Sand pebbel images
I9 0 . 247 0 . 253 0 . 289

I 10 0 . 254 0 . 257 0 . 318

Table 2. ERLBP applied to sandpebbels crack and beach images

Figure 4. Beach images

Figure 6. ELBP on sandpebbels and crack images

Figure 5. Crack images

IMAGE SPELBP CELBP BELBP


I1 0.039 0.041 0.026

I2 0.0431 0.044 0.031

I3 0.044 0.045 0.041 Figure 7. ELBP on sandpebbels and beach images


I4 0.048 0.048 0.042

I5 0.048 0.059 0.045

I6 0.05 0.066 0.048

I7 0.054 0.07 0.063

I8 0.054 0.076 0.063

I9 0.06 0.081 0.066

I 10 0.065 0.113 0.067

Table 1. ELBP applied to sandpebbels crack and beach images

Figure 8. ERLBP on sandpebbels and crack images

i-manager’s Journal on Pattern Recognition, Vol. 1 l


No. 1 l
March - May 2014 27
ARTICLE
RESEARCH PAPERS

Man Cybernetics, 3 : P P. 6 1 0 - 6 2 1 . DOI:


10.1109/TSMC.1973.4309314
[5]. J. Daugman. (1985). “Uncertainty Relation for
Resolution in Space, Spatial Frequency and Orientation
Optimised by Two-Dimensional Visual Cortical Filters”,
Journal of the Optical Society of America, 2, pp.1160-
Figure 9. ERLBP on sandpebbels and beach images 1169.

Conclusion [6]. T. Ojala, and M. Pietikainen. (2004). “Texture


Classification”. Machine Vision and Media Processing
The LBP operator has been theoretically simple, yet a very
Unit University of Oulu, Finland.
powerful method of analyzing textures. ELBP indicate
overlapping of two different texture features and may [7]. T. Ojala, M. Pietikainen and D. Harwood. (1996). “A

exhibit better results with more samples. ERLBP clearly comparative study of texture measures with classification

distinguish different images without any overlapping and based on featured distributions”. Pattern. Recognition.

the performance is independent of the number of 29: pp.51-59. DOI: 10.1016/0031-3203(95)00067-4.

samples. ERLBP is computationally more efficient as the [8]. O.Rama Devi and Dr. E. V. Prasad, Dr. L. S. S. Reddy.
code matrix never exceeds 8x2 sizes as against 8x9. (2013). “Robust Rule Based Local Binary Pattern Method
Finally ERLBP undoubtedly be a better classifier. for Texture Classification and Analysis”. IIJETTCS. pp.274-

References 277.

[1]. A. Rosenfeld and J. Weszka. (1980). “Picture [9]. R. Haralick. (1979). “Statistical and Structural

Recognition”. Digital Pattern Recognition, K. Fu (Ed.), Approaches to Texture”. Proc. IEEE, 67, 5, pp.786-804.

Springer-Verlag, pp.135-166. [10]. T. Ojala, M. Pietikainen and T. Maenpaa. (2002).

[2]. D. Huang, Y. Wang, and Y. Wang. (2007). “A robust “Multiresolution gray-scale and rotation invariant texture

method for near infrared face recognition based on classification with local binary patterns,” IEEE Transaction

extended local binary pattern,” in Proc. Int. Symp. Vis. on Pattern Analysis Machine Intelligence. Vol. 24, no. 7,

Comput., pp.437–446. pp.971–987.

[3]. G. Cross and A. Jain. (1983). “Markov Random Field [11]. J. S. Weszka, C. R. Dyer and A. Rosenfeld. (1976). “A

Texture Models”, IEEE Trans. Pattern Analysis and Machine comparative study of texture measures for terrain

Intelligence, 5, 1, pp.25-39. classification”. IEEE Transaction on System Man


C y b e r n e t i c s . 6 : p p . 2 6 9 - 2 8 5 .
[4]. R.M Haralick, K. Shanmugam and I. Dinstein. (1973).
DOI:10.1109/TSMC.1976.5408 777.
“Textural features for image classification”. IEEE Trans. Syst.

28 i-manager’s Journal on Pattern Recognition, Vol. 1 l


No. 1 l
March - May 2014
ARTICLE
RESEARCH PAPERS

ABOUT THE AUTHORS


O. Rama Devi is working as an Assistant Professor in P. V. P. Siddhartha Institute of Technology, Kanuru. She received her B. Tech
degree in CSE from K. L. College of Engineering, Guntur and M. Tech degree in Computer Science from K. L. College of
Engineering, Guntur. Currently, she is pursuing her PhD at JNTUK, Kakinada. She has 6 years of teaching experience. She is a
member of CSI professional societies. She has published 4 research papers in various national and international Journals and
attended 10 national and international Conferences. Her areas of interests are image processing, computer networks and
pattern recognition.

Dr. E. V. Prasad is the Director of LBR college of Engineering, Mylavaram. He was retired as a rector, JNTU Kakinada. He held
different positions Principal, Director, Registrar and act as a BOS Chairman. He has supervised 7 PhD students and 8 students
working for their degree. He is the Co-author of six books and published hundred papers in national and International journals
and conferences. His research interests include Data Mining, and Information Security.

Dr. L. S. S. Reddy is currently the Pro.VC of KL University, Vaddeswaram, Andhra Pradesh. He received his B. Tech degree in
Electronics and communication engineering from J. N. T. University, Hyderabad; M. Phil degree in Computer Science from
Central University, Hyderabad, and PhD from BITS, PILANI. He has more than 20 years of teaching experience. He has published
fifty-two research papers in various national and international Journals and more than twenty research papers in various national
and international Conferences. He is also the Chairman Board of study in Nagarjuna University. He has guided four PhD scholars
and Chairman of CSI Chapter, Vijayawada. He is a member of various professional societies like IEEE, ISTE, IETE, IE and CSI.

i-manager’s Journal on Pattern Recognition, Vol. 1 l


No. 1 l
March - May 2014 29

You might also like