Efficient Texture Classification Using a Kohonen Clustering Network and the LNLBP Attributes

(1)

Efficient Texture Classification Using a Kohonen

Clustering Network and the LNLBP Attributes

Abstract–In this paper, a Kohonen clustering network is proposed for efficient texture classification. Our goal is to be able to determine with accuracy different classes of similar and superposed textures. To this end, we introduce a new concept of local binary patterns called large neighborhoods local binary pattern (LNLBP), for discriminative network classification. The processed pixel to be classified considers window of large neighborhoods perversely to classic techniques that consider small sized windows. In addition, the use of characterizing parameters and a study for optimal windows size selection are proposed. A database composed by image holding similar textures patterns is used. The proposed approach generates classification results with high accuracy and reliability. A comparison study is conducted and proved that this approach is more efficient than many recent published methods.

Keywords – Texture Classification, Kohonen Networks, Large Neighborhoods Local Binary Patterns (LNLBP), LBP.

I. I

NTRODUCTION

Texture analysis has gained an important role in many applications in computer vision, image retrieval and motion analysis. In case of superposed or adjacent textures in the same image a main problem lies in identifying with accuracy the different regions with near content, and borders between them.

In the literature, a large number of approaches have been proposed. Different advanced techniques in image classification using different techniques such as: artificial neural networks (ANN), support vector machines (SVM), fuzzy approach, fuzzy support vector machines (FSVM) and genetic algorithms with neural networks are being developed [1]. Ojala et al. [2] proposed the uniformed local binary patterns (LBP) approach, which was extended by huang, li and wang by computing the derivative-based local binary patterns and applied it to the face alignment [3]. Still with LBP, S. liao and A.C.S. chung introduced the concept of advanced local binary patterns (ALBP) with the Aura Matrix measure as the second layer to analyze texture [4]. Also, M.T. aloui and sbihi proposed in [5] a method for unsupervised texture classification, based on both kohonen maps and mathematical morphology. Whereas recently, L.Tlig, presented in [7] a new fuzzy segmentation approach based on S-FCM type2 using LBP-GCO feature improving the classification results and decreasing considerably the error rate. In the last section of this paper a comparative study between the results of the proposed approach and those of [7] is conducted. The proposed approach is divided in two essential steps: 1. A Pre-processing step: it consists in finding the optimal

window size belonging to each different texture in the image able to characterize it with high accuracy. The

inputs of the kohonen clustering network are the histograms of the LNLBP attributes.

2. Training and classifying step: Training the network, collecting the decision and comparing the maps outputs with the desired Patterns using different evaluation criteria.

The paper is organized as follows: section2 describes the kohonen clustering network. Section3 discusses the proposed LBP method and details the various pre-processing steps. Section4 deals about the experimental results and the comparative study.

II. K

OHONEN

C

LUSTERING

N

ETWORK

A. General Description

The basic idea behind the kohonen network (or “self organizing map” (SOM)) is to set up a structure of interconnected processing units (“neurons”) which compete for the signal. This map structure is composed by two layers: the input layer with N neural units and the competitive layer (output layer) with M neural. Each interconnection between the two layers has a weight W as presented by the following figure.

Fig.1. Kohonen feature map

When an input is presented to the network, the neural unit whose weight vector is the closest to this observations wins the competition, and according to this situation the output of the winner is then equal to 1 while the outputs of all the other output units are set to 0 [5], as presented by the following equation:

 

  

 else

M M t W t X d t W t X d if

Y_M Q M Q M

0

)) ( ), ( ( )) ( ), ( (

1 ₁

1 ₍₁₎

Where:

Q

X is the input vector, W_M is the weight vector for the output unit M, d(X_Q(t),W_M₁(t)):the Euclidean distance between the observation X_Q(t) and the weight vector

1

M

(2)

International Journal of Electronics Communication and Computer Engineering Volume 4, Issue 3, ISSN (Online): 2249–071X, ISSN (Print): 2278–4209

The weight vector of the winning unit, noted M2, and its neighbors M are modified according to equations [5]:

2 2 2 2 2 )) ( , ( ) 1 ( ) ( )) ( , ( )] 1 ( ) ( ).[ , ( ). ( ) 1 ( ) ( )] 1 ( ) ( ).[ ( ) 1 ( ) ( M M and t r M V M if t W t W t r M V M if t W t X t M h t a t W t W M M if t W t X t a t W t W M M M Q M M M Q M M                    (2)

M2is the winning unit defined by the following equation: ))] ( ), ( ( [ min

2 Arg d X t W t

M _Q _M

m

 (3)

r(t) is the interaction radius which depends on the number of the iterations t. a(t) is the learning coefficient at the time t. V(M,r) is the neighborhood of a neural unit m with a radius r, defined by:





M M M M dU U r



r M

V( , ) 10, , 1 / ( M, M1) (4)

Where ( , )

1

M M U

U

d denotes the Euclidean distance betweenU and_M

1

M

U . h(M2,t) is the interaction function that depends on the proximity radius r(t) defined by:

) ) ( 2 ) , ( exp( ) , ( 2 2 2 1 t r U U d t M

h   M M (5)

B. The Training Algorithm

The basic training algorithm is detailed as follows: 1. Select an “object” from the training set.

2. Finding the minimum distance between W_i_,_j and the training data (the node which is closest to the selected data).

3. Adjust the weight vectors of the closest node and its neighbors such a way that the W_i,_j move towards the

training data.

4. Repeat from step1 for a fixed number of iterations.

III. T

EXTURES

P

RE

-P

ROCESSING

A. Size of the training windows

For a reliable classification it is very important to determine an adequate amount of information that is able to characterize texture with the highest accuracy. If the window gathering the discriminative information is small, the classifier is enabling to distinguish different classes, generate high error classification rate. Whereas a large window will characterize different types of similar textures, which introduces a high confusion in the output decision and high errors in the obtained results. In this proposed approach a novel strategy is presented aiming to find out the optimal window size selected from the texture and processed to determine the features able to characterize the desired data. We propose to compute the coefficient of variation defined by equation 6 that describes the normalized variation of a chosen texture for different windows size. The windows size increase with a step of (3×3) pixels without overlap to the maximal size of the texture noted (N×M). This generates a curve characterized by a transitory phase and a stable phase. This last phase indicates that increasing the windows size doesn’t lead to more accuracy for the texture characterizing. Ones the function is stable we extract the

first window size corresponding to last tangent before the curve becomes linear as presented by figures (2-3).

 



CV (6)

Where the standard deviation and μ the mean:



 

 

 Ni

i j N j j i j i x N N ₁ 2 1 ) ) , ( ( 1 _  (7)



  

 Ni

i j N j j i j i x N

N ₁ ₁ (, )

1

 (8)

The optimal size corresponds to this for which the coefficient of variation presents a constant behavior. This stable zone means that the increase of the window size doesn’t improve the texture characterization. In this zone, more the size of the window increase more the number of attributes and the processing time increase without better results.

Fig.2. Coefficient of variation corresponding to the texture shown in the upper left corner of the figure

Fig.3. Coefficient of variation corresponding to the texture shown in the upper left corner of the figure Fig.2 and 3 show the evolution of the coefficient of variation in relation with the window size (illustrated in blue) used for the calculation of the textures parameters. The window size is extracted from curves where n is the window index and its size is



lk

 

 2n1

 

 2n1



. Once the allure of the coefficient of variation noted CV(n) is determined, we compute its derivate to find the optimal value of the index called n_optas presented by the equation 9. The segment D (illustrated in black) is the tangent to the fitting curve (illustrated in red) at the index 12 which represents the optimal window size for these two textures. D is defined by the following equation:

2 Arg d X t W t

M _Q _M

m

 (3)





M M M M dU U r



r M

V( , ) 10, , 1 / ( M, M1) (4)

Where ( , )

1

M M U

U

1

M

) ) ( 2 ) , ( exp( ) , ( 2 2 2 1 t r U U d t M

h   M M (5)

B. The Training Algorithm

training data.

III. T

EXTURES

P

RE

-P

ROCESSING

A. Size of the training windows

 



CV (6)



 

 

 Ni

i j N j j i j i x N N ₁ 2 1 ) ) , ( ( 1 _  (7)



  

 Ni

i j N j j i j i x N

N ₁ ₁ (, )

1

 (8)

Fig.3. Coefficient of variation corresponding to the texture shown in the upper left corner of the figure Fig.2 and 3 show the evolution of the coefficient of variation in relation with the window size (illustrated in blue) used for the calculation of the textures parameters. The window size is extracted from curves where n is the window index and its size is



lk

 

 2n1

 

 2n1



. Once the allure of the coefficient of variation noted CV(n) is determined, we compute its derivate to find the optimal value of the index called n_optas presented by the equation 9. The segment D (illustrated in black) is the tangent to the fitting curve (illustrated in red) at the index 12 which represents the optimal window size for these two textures. D is defined by the following equation:

2 Arg d X t W t

M _Q _M

m

 (3)





M M M M dU U r



r M

V( , ) 10, , 1 / ( M, M1) (4)

Where ( , )

1

M M U

U

1

M

) ) ( 2 ) , ( exp( ) , ( 2 2 2 1 t r U U d t M

h   M M (5)

B. The Training Algorithm

training data.

III. T

EXTURES

P

RE

-P

ROCESSING

A. Size of the training windows

 



CV (6)



 

 

 Ni

i j N j j i j i x N N ₁ 2 1 ) ) , ( ( 1 _  (7)



  

 Ni

i j N j j i j i x N

N ₁ ₁ (, )

1

 (8)

(3)

( ) )(

(

' n_opt n n_opt CV n_opt CV

D    (9)

So, for the classification of these images we will use windows having size (25×25).

B. The Local Binary Pattern

The LBP operator [6] is a characterizing pattern applied to discriminate texture. It can be defined as:







 

 1

0

, ( ) 2

N i

i C i R

i signU U

LBP (10)

Where N is the total number of neighbors, Uc is the

intensity of the center pixel, Ui is the intensity of the

neighbor i and sign(x) is the step function.

Considering neighborhood of (3×3) pixels, the LBP value for such a neighborhood is estimated as follows:

1. The original (3×3) neighborhood (Fig4.a) is thresholded in binary levels (0 and 1) by using the function “sign(x)” (i.e. sign(x)=1 when x≥0 and sign(x)=0 with respect to the value of the central pixel). 2. The values of the thresholded pixels are multiplied by

the binomial weights (Fig4.c)

3. The values of the eight pixels (Fig4.d) are summed to obtain a single value for the corresponding pattern.

Fig.4. An example of LBP value estimation

C. The proposed (LNLBP) method

In this paper, we propose a new approach for LBP computing presented in four steps as follows:

1. The original image with size (X×Y) is surrounded by 9 null matrices with size (X+2, Y+2). Each matrix M is_i oriented in 8 directions, presented as follows (in fig.5): 2. All the elements of each matrix M are thresholded toi

be transformed in binary levels (0 and 1). This is based on a directional comparison between the extended matrixM and the original matrix M as presented byi

the following equation and (Fig6):

0 ) , ( 1 ) , ( )

, ( ) ,

(x y M x y thenM x y  elseM x y  M

if _i _i _i

3. Each obtained M matrix is multiplied by_i 2(i1) . where I represents the matrix order or index between the eight matrices (Fig6).

4. The obtained matrices are summed to obtain a new matrix M that characterizes the processed pattern. The matrix kernel produced through these steps is considered. Its histogram is computed and used as discriminative parameter to model the texture. It represents the input of the kohonen clustering network.

Fig.5. Matrices with their various directions, centered in M

Fig.6. The proposed method (LNLBP)

IV. R

ESULTS

The experimental results are divided into two essential parts: the first one illustrates the results of simulations and the second propose a comparison study between our proposed method and recent approach in the literature for texture classification [7].

A. Experimental Results

In the conducted experiments, we applied the proposed algorithm to segment four textures. For each image, we computed the optimal window size. Then, we applied three different evaluations criteria:

1. Region differencing: several measures operate by computing the degree of overlap between clusters [8]. 2. Accuracy measure: it is used to estimate the correct

and false classification for each output [7] (equation 11).

( ) )(

(

D    (9)

B. The Local Binary Pattern







 

 1

0

, ( ) 2

N i

i C i R

i signU U

LBP (10)

C. The proposed (LNLBP) method

0 ) , ( 1 ) , ( )

, ( ) ,

if _i _i _i

IV. R

ESULTS

A. Experimental Results

and false classification for each output [7] (equation 11).

( ) )(

(

D    (9)

B. The Local Binary Pattern







 

 1

0

, ( ) 2

N i

i C i R

i signU U

LBP (10)

C. The proposed (LNLBP) method

0 ) , ( 1 ) , ( )

, ( ) ,

if _i _i _i

IV. R

ESULTS

A. Experimental Results

(4)





 N

i TP

CP Accuracy

1

(11) Where:

CP is the number of correct segmented pixels in ith texture.

TP is the total number of pixels in ithtexture.

3. Vinet criteria: it calculates a dissimilarity measure between the real and desired classification. The maximum overlap between the two is computed. In this case, the two classes are represented by the desired segmentation (referred by ‘b’ in the previous figures) and the Kohonen segmentation (referred by ‘c’ in the previous figures). This method assumes that two classes are matched if they have a common maximum pixel set. For any pair of regions (Vi, Rj), their recovery is given as follows:

) ( _i _j

ij cardV R

t   (12)

We supposed that H is the number of obtained pairs and C1…CH are the recoveries of each of these pairs; The measurement of Vinet is then defined by the next equation (equation 13):



  

 H

K

k C Y X Vinet

1

) ( 1

1 (13)

Where (X×Y) represents the size of the original image. In our experiments, we applied the proposed algorithm to segment four textures:

Table I: Optimal Sample Size for the 4 Textures Figures Optimal Window

Size

Number Windows For The Network Training

Fig7 (35×35) Only 1 window

(a)

(b) (c)

Fig.7. (a) Original image; (b) desired segmentation; (c) segmentation results using the proposed method

To train the network in order to classify the textures, only one window pattern is introduces. This fact is very important since it decreases hugely the processing time when compared with methods based on neural network classification

(a)

(b) (c)

(a)

(b) (c)





 N

i TP

CP Accuracy

1

(11) Where:

) ( _i _j

ij cardV R

t   (12)



  

 H

K

k C Y X Vinet

1

) ( 1

1 (13)

Size

(a)

(b) (c)

(a)

(b) (c)

(a)

(b) (c)





 N

i TP

CP Accuracy

1

(11) Where:

) ( _i _j

ij cardV R

t   (12)



  

 H

K

k C Y X Vinet

1

) ( 1

1 (13)

Size

(a)

(b) (c)

(a)

(b) (c)

(a)

(b) (c)

(5)

(b) (c)

Fig.10. (a) Original image; (b) desired segmentation; (c) segmentation results using the proposed method The following table shows the evaluation results of the classified images in figures (7, 8, 9 and10). Figures (a), (b) and (c) represent respectively the original texture, the desired classification and the classification obtained by the proposed method.

Table II: Evaluation of texture classification Evaluation

Method

Fig7 Fig8 Fig9 Fig10

Region differencing

96.19 % 97.92 % 98.28 % 98.96 %

Accuracy measure

97.44 % 97.65 % 98.27 % 98.80 %

Vinet criteria

86.19 % 98.47 % 98.58 % 99.66

B. Comparative Study

In order to test the segmentation performance of our proposed method, we compared our approach with two other techniques proposed recently in [7]: the FCM type2 and S-FCM type2. Each segmentation technique has been applied on a set of 16 composed textures shown by Fig.11. The goal is to have the lowest error rate presented by the following equation:

100

(%) 

Ap Mp rate

error (14)

With Mp and Ap represent respectively the numbers of misclassified pixels and the entire image pixels.

Fig.11. 16 images presenting mixtures of two textures of (128×128) pixels [7]

Table III: Error rate of three fuzzy segmentation approaches applied on the set of 16 textured images

illustrated in Fig.11

Images FCM

type2

S-FCM type2

LNLBP method

Im1 7.18 % 5.93 % 2.13 %

Im2 7.96 % 2.81 % 2.09 %

Im3 8.75 % 6.56 % 2.93 %

Im4 11.40 % 8.28 % 5.60 %

Im5 5.62 % 2.96 % 1.61 %

Im6 4.06 % 2.03 % 1.59 %

Im7 2.96 % 2.03 % 1.91 %

Im8 6.09 % 5.46 % 1.19 %

Im9 11.87 % 7.96 % 2.27 %

Im10 16.87 % 11.40 % 3.64 %

Im11 8.28 % 7.81 % 2.06 %

Im12 13.59 % 7.81 % 3.66 %

Im13 18.43 % 10.46 % 2.09 %

Im14 5 % 3.43 % 1.50 %

Im15 13.90 % 8.59 % 2.55 %

Im16 10 % 7.50 % 2.85 %

Fig.12. Comparative study between the LNLBP and recent techniques in use

0 2 4 6 8 10 12 14 16 18

0 2 4 6 8 10 12 14 16 18 20

tes t im ages

er

ro

r

ra

te

%

(6)

This graph shows that the proposed LNLBP method (with yellow) generates the best results and the less classification error when compared with recent techniques in use.

V. C

ONCLUSION

In this PAPER, a new approach called LNLBP is presented. The proposed method combines a new concept of LBP called (LNLBP) with a kohonen clustering network. The texture is characterized by it coefficient of averaged variation. Exploiting large LBP neighborhood and an accurate selection of the optimal window size of coefficient of variation is found to be very efficient for a reliable classification.

A comparison study is conducted to highlight the efficiency of the proposed method face to recent methods in the literature.

R

EFERENCES

[1] M. Seethe I. V. Muralikrishna B. L. Deekshatulu B. L.

Malleswari and P. Hegde, “artificial neural networks and other

methods of image classification,”,JATIT, pp. 1039-1053, 2005-2008

[2] M. pietikainen T. ojala and T. maenpaa, “Multiresolution gray -scale and rotation invariant texture classification with local

binary patterns,” PAMI, vol. 24, no.7, pp. 971-987, 2002 [3] S. Z. Li X. Huang and Y. Wang, “shape localization based on

statistical method using extended local binary pattern,” in IEEE

proc, conf. Image and graphics, 2004, pp.184-187

[4] Shu liao and Albert C. S. Chung, “texture classification by using

advanced local binary patterns and spatial distribution of

dominant patterns,” in IEEE, conf. ICASSP, pp. 221-224, 2007 [5] Mohammed talibi-alaoui and abderrahmane sbihi, “fractal

features classification for texture image using neural network

and mathematical morphology,”, WCE, VOL. I, 2012

[6] M. A. Savelonas D. K. Iakovidis D. E. Maroulis, “An LBP-based

active contour algorithm for unsupervised texture segmentation,” in IEEE, conf.ICPR’06, 2006

[7] Lotfi tlig, Mounir sayadi, Farhat fnaiech, “a new fuzzy

segmentation approach based on S-FCM type2 using LBP-GCO features,” Elsevier, conf. Image communication, vol. 27, pp.

694-708, 2012

[8] R. unnikrishnan C. pantofaru M. Hebert, “a measure for objective evaluation of image segmentation algorithms,” in

IEEE, 2005

A

UTHOR

’

S

P

ROFILE

M. Bahri

was born in Tunisia in 07 June 1989. He received the M.S. degree from National higher school of engineer (ENSIT) in Tunis, Tunisia, in 2013.

He is author of 3 conference papers and currently working towards the Ph.D. Degree in the division of Electronic and digital technologies at ENSIT. His current research focuses on texture and image analysis.

H. Seddik

was born in Tunisia in 15 October 1970, he has obtained the electromechanical engineer degree from the national higher school of engineer (ENSIT) in Tunis, Tunisia in 1995 and followed by the master degree in “signal processing: speaker recognition”

and the thesis degree in image processing

“watermarking using non conventional transformations”.

He is currently an assistant professor in ENSIT and works in the field since 17 years. He has authored more than 10 papers and book chapters in international journals and about 47 conference papers. His domain of interest is: Audio-image and video processing applied in filtering, encryption and watermarking.

Dr Seddik belongs to the CEREP research unit and supervises actually 04 thesis and 08 masters in the field.

A. Selmani

was born in Tunisia in 20 February 1988. She received the M.S. degree from National higher school of engineer (ENSIT) in Tunis, Tunisia, in 2012.

Efficient Texture Classification Using a Kohonen Clustering Network and the LNLBP Attributes