Haralick feature extraction from LBP images for color texture classification

Transcription

1 Image Processing Theory, Tools & Applications aralick eature extraction rom LBP images or color texture classiication Alice Porebski,2, icolas Vandenbroucke,2 and Ludovic Macaire 2 École d Ingénieurs du Pas-de-Calais (EIPC) Département Automatique - Campus de la Malassise Longuenesse Cedex - FRACE alice.porebski@eipc.r, nicolas.vandenbroucke@eipc.r 2 Laboratoire LAGIS - UMR CRS 846 Université des Sciences et Technologies de Lille Cité Scientiique - Bâtiment P Villeneuve d Ascq - FRACE ludovic.macaire@univ-lille.r Abstract In this paper, we present a new approach or color texture classiication by use o aralick eatures extracted rom co-occurrence matrices computed rom Local Binary Pattern (LBP) images. These LBP images, which are dierent rom the color LBP initially proposed by Mäenpää andpietikäinen, are extracted rom color texture images, which are coded in 28 dierent color spaces. An iterative procedure then selects among the extracted eatures, those which discriminate the textures, in order to build a low dimensional eature space. Experimental results, achieved with the BarkTex database, show the interest o this method with which a satisying rate o wellclassiied images (85.6%) is obtained, with a -dimensional eature space. Keywords Color texture classiication, Feature extraction, LBP images. I. ITRODUCTIO Color texture classiication is a major ield o development or various vision applications, and particularly or the industrial quality control where color textures have to be characterized in order to detect deects on color texture areas and sort the products into dierent categories []. Many authors have shown that the use o color improves the characterization o color textures and consequently the results o texture classiication [2], [3], [4], [5]. That is why many relevant texture descriptors, initially deined or grey images, have been extended to color and used to classiy color textures, like Markov random ields [6], wavelet transorm [7], [8], [9], co-occurrence matrices [3] or Local Binary Patterns (LBP). LBP, which have initially been proposed in 996 by Ojala to describe the textures present in grey level images [], have then been extended to color by Mäenpää and Pietikäinen [2], []. The use o the LBP images in order to characterize color textures is very expensive, since it requires 9 LBP images deduced rom the original color image. Indeed, LBP images are based on a scalar analysis o colors. In this paper, we propose a new color LBP image, based on a vectorial analysis o colors. This new approach provides only one single color LBP image which characterizes the color texture. In order to classiy color textures, we propose to ollow a classical approach, deined in section III : we compute the cooccurrence matrix o this new color LBP image and extract well-known aralick eatures rom this matrix. Futhermore, the analysis o the color properties is not restricted to the acquisition color space (R, G, B) (see section II) and there exists a lot o color spaces which respect dierent properties [2]. one o them is adapted to the classiication o all kinds o color textures [3]. That is why we propose to select the texture eatures rom color texture images coded in dierent color spaces (see section IV) [4]. The selection o the aralick eatures computed in dierent color spaces signiicantly improves the classiication quality and also allows to work with a low-dimensional eature space [3]. This is important within the ramework o industrial quality control to decrease the processing time. This discriminating low-dimensional eature space is built by using the iterative selection procedure described in the section IV-B. Experimental results, achieved with the BarkTex benchmark database, show the interest o this approach in the last section [5]. II. COLOR REPRESETATIO A. Color space and texture analysis Color analysis is not restricted to the (R, G, B) color space and there exists a large number o color spaces which respect dierent properties [2]. These color spaces can be classiied into our amilies : the primary color spaces, the luminancechrominance color spaces, the perceptual color spaces and the independent color component spaces (see Fig. ). Figure shows that these amilies can be divided in subamilies too. Furthermore, Palm, Drimbarean and Chindaro have compared the perormances o color texture classiication reached by using color texture eatures extracted rom images whose pixel color is represented in dierent color spaces [3], [4], /8/$ IEEE

2 (R, G, B) (X, Y, Z) (x, y, z) 2 (r, g, b) Primary spaces Real and artiicial primary spaces ormalized 2 coordinate spaces Antagonist 3 spaces Television 4 spaces Perceptually 5 uniorm spaces Other luminancechrominance spaces 6 Polar 7 coordinate spaces Perceptual 8 coordinate spaces Fig.. (A, C,C2) 3 (bw, rg, by) (Y,I,Q ) 4 (Y,U,V ) (L,a,b ) 5 (L,u,v ) (L, U, V ) 6 (Y,x,y) (I,r,g) Luminancechrominance spaces (I,I2,I3) Independent component spaces Color space amilies. (A,C CC2,h CC2 ) 7 (bw,c rgby,h rgby ) (Y,C IQ,h IQ ) (Y,C UV,h UV ) (L,C ab,h ab) (L,C uv,huv) (L, C UV,h UV ) (I,C I2I3,h I2I3 ) (L,Suv,h uv) 8 (I,S,3) (I5,S4,2) (I4,S3,2) (I,S2,) (I,S,) Perceptual spaces [5]. The synthesis o these works does not allow to conclude on the deinition o a single color space adapted to color texture analysis. In order to take into account the properties o dierent color spaces, Chindaro proposes to merge dierent classiiers where the images are coded in dierent color spaces [5]. Likewise, Vandenbroucke proposes to select local statistical eatures, which are computed rom dierent color components [6]. In this paper, we propose also to associate several color spaces in order to characterize the textures by extracting the color texture eatures rom color images coded in each o the S =28color spaces o Fig.. B. Color order relation In our approach, the color ranks o pixels are compared to compute the LBP images (see section III-A.2). The color o pixels is represented by a vector, but as there does not exist a total order between vectors, we need to consider a partial order relation [8]. We choose to use or its simplicity, the partial order relation used by vector median ilters and deined as ollows : For each color space S =(C,C 2,C 3 ) o Fig., the color a =[C a C2 a C3 a ] precedes the color b =[C b C2 b C3], b with respect to the origin point (,, ), i (C a)2 +(C2 a)2 +(C3 a)2 (C b)2 +(C2 b)2 +(C3 b)2. This rule is based on the comparison o the norms o the two color vectors. Ater having presented the S = 28 color spaces used to characterize more eectively the textures, and the order relation required to compare two colors, we explain in the next section the computation o the color texture eatures. III. COLOR TEXTURE FEATURES In this section, we irstly explain how the LBP images are computed rom the original color texture images. Then, we detail the computation o the co-occurrence matrices which are extracted rom these LBP images, and inally we present aralick eatures which are used to reduce the large amount o inormation o the co-occurrence matrices, while preserving their relevance. A. Local binary pattern images ) Scalar LBP images: Local Binary Patterns (LBP) have initially been proposed in 996 by Ojala to describe the textures present in grey level images []. These texture descriptors are very interesting because they are particularly well-adapted to real-time quality control applications as they are both ast and easy to implement [9]. They have then been extended to color by Mäenpää and Pietikäinen and used in several color texture classiication problems [2], []. These color texture descriptors are deined as ollows : Let C k and C k, be two o the three color components o the color space S = (C,C 2,C 3 ) (k, k {, 2, 3}) and LBP C k,c k [P ], be the LBP which represents the local pattern in the neighborhood o the pixel P, or the components C k and C k : irst, the color component C k o each pixel P o the neighborhood is thresholded into two levels ( and ) by using the color component C k (P ) o the considered pixel P as threshold T : i C k (P ) C k (P ), then C k (P )=, else C k (P )=. the result o each thresholding is then coded thanks to a weight mask : the weighted values are inally summed in order to obtain the value o the LBP LBP C k,c k [P ]. Figure 2 illustrates the computation steps achieved to obtain the LBP images LBP C,C C,C2 8 [I], LBP8 [I] and LBP C2,C 8 [I] extracted rom the original image I, whose the pixel color is represented by the cell, and where the neighborhood here considered to compute these LBP images is the 8-neighborhood, denoted 8 (see Fig. 4). For example, or the computation o LBP C,C2 8 [P ], P being represented by the cell 2 C C 2 C 3 in Fig. 2, the color component C 2 o each o the 8 neighboring pixels is compared with the color component C (P ) = 2 = T. For the neighboring pixel located down let, or example, the

3 Original color texture image I Original color texture image I C,C C,C2 C2,C Thresholding Weight mask Thresholding T T T = = = T = Weight mask Weighting Weighting Sum Sum LBP C,C 8 [I] LBP C,C 2 8 [I] LBP C 2,C 8 [I] LBP image LBP C,C 2,C 3 8 [I] Fig. 2. The dierent steps to obtain the LBP images LBP C,C 8 [I], LBP C,C 2 8 [I] and LBP C 2,C 8 [I] extracted rom the original image I. result o the thresholding is (2 T ). This result is then weighted by 32. Ater having weighted the 8 thresholding values, we sum them and obtain LBP C,C2 8 [P ] = = 36. For a given neighborhood, a color image I coded in the (C,C 2,C 3 ) color space is characterized by the 9 ollowing color LBP images : LBP C,C [I], LBP C2,C2 [I], LBP C3,C3 [I], LBP C,C2 I], LBP C2,C [I], LBP C,C3 [I], LBP C3,C [I], LBP C2,C3 [I] et LBP C3,C2 [I]. Pietikäinen and Mäenpää propose to compute the histogram rom each o these 9 LBP images, and to concatenate these histograms into a single one to characterize the color textures. They then use a log-likelihood dissimilarity measure to classiy the color texture images with this LBP distribution [2], []. 2) Vectorial LBP images: Our approach diers rom this initial deinition. Indeed, instead o comparing the color components o pixels, we compare their color rank thanks to the order relation deined in the section II-B. Fig. 3. The dierent steps done to obtain the color LBP image LBP C,C 2,C 3 8 [I] extracted rom the original image I. This color texture descriptor is deined as ollows : Let LBP S [P ], be the LBP image which represents the local pattern in the neighborhood o the pixel P coded in the color space S : For each pixel P, we irstly compare the color C(P )= [C (P ) C 2 (P ) C 3 (P )] o this pixel with the color C(P )=[C (P ) C 2 (P ) C 3 (P )] o each neighboring pixel P, thanks to the color order relation : i (C (P )) 2 +(C 2 (P )) 2 +(C 3 (P )) 2 (C (P )) 2 +(C 2 (P )) 2 +(C 3 (P )) 2, then C(P )=, else C(P )=. the result o each thresholding is then coded thanks to the weight mask proposed by Mäenpää and Pietikäinen [2], [], the weighted values are inally summed in order to obtain the value o the color LBP LBP S [P ]. Figure 3 illustrates the computation steps done to obtain the color LBP image LBP (C,C2,C3) 8 [I] rom the original image I whose the pixel color is represented by the cell C C 2 C 3.

4 For example, or the pixel P, P being represented by the cell in Fig. 3, the color o each o its 8 neighbors 2 is compared with the threshold T = (2) 2 + () 2 + () 2 thanks to the ollowing relations : ) () 2 + () 2 + () 2 <T 2) () 2 + () 2 + (2) 2 T 3) () 2 + (2) 2 + () 2 T 4) (2) 2 + () 2 + () 2 T 5) () 2 + () 2 + (2) 2 T 6) () 2 + () 2 + () 2 <T 7) (5) 2 + (2) 2 + (5) 2 T 8) () 2 + () 2 + () 2 <T The pixels whose color precedes the pixel P ones are labeled. Otherwise they are labeled (see Fig. 3). The result o each thresholding is then coded thanks to the weight mask and the weighted values are inally summed to obtain the value o LBP (C,C2,C3) 8-neighborhood [P ] : LBP (C,C2,C3) 8 [P ]= =82. The strong point o our approach is that it allows to characterize color textures only with one color LBP image, contrary to Mäenpää and Pietikäinen s deinition where 9 LBP images are extracted rom the original image to characterize the color textures. Otherwise this deinition o color LBP images allows to emphasize the local color variations, as we will see in the section V-B. Thereore it is interesting to extract rom these LBP images, texture eatures which measure the grey scale distribution and consider the spatial interactions between pixels, instead o extracting histograms, as Mäenpää and Pietikäinen do. We thus choose to use aralick eatures computed rom co-occurrence matrices to test the eectiveness o the LBP images or color texture classiication. As they measure the local interaction between pixels, the cooccurrence matrices are sensitive to signiicant dierences o spatial resolution and image size. To decrease this sensitivity, it is necessary to normalize these matrices by the total cooccurrence number i= j= M [I](i, j), where is the quantization level number. The normalized color cooccurrence matrix m [I] is deined by : m [I] = i= M [I] j= M [I](i, j). Dierent neighborhoods can be considered to compute the co-occurrence matrices. The irst element to be considered is the shape o the neighborhood. Figure 4 shows dierent 3x3 neighborhood shapes. 8-neighborhood 4-neighborhood 4-neighborhood 2 2-neighborhood 2-neighborhood 2-neighborhood 2-neighborhood direction 9 direction 45 direction 35 direction Fig. 4. Example o dierent 3x3 neighborhood shapes in which neighboring pixels are labeled as gray. The choice o the neighborhood shape depending on the analysed textures [2], we will see in section V-B that the 8- neighborhood is the best-adapted or our application since it takes into account all directions. The second element o the neighborhood to be considered is the distance d between the considered pixel P and its neighbors. Figure 5 illustrates the 8-neighborhood, or a given distance d. B. aralick eatures extracted rom co-occurrence matrices Co-occurrence matrices, introduced by aralick [2], are statistical descriptors which both measure the grey scale distribution in an image and consider the spatial interactions between pixels. These texture descriptors are deined as ollows : Let M [I], the co-occurrence matrix which measures the spatial interactions between the pixels o the image I. The cell M [I](i, j) o this matrix contains the number o times that a pixel P whose grey level G(P ) is equal to i, isthe neighbor o a pixel Q whose grey level G(Q) is equal to j, according to the neighborhood. d Fig. 5. Illustration o the neighborhood used to compute the co-occurrence matrices, or a given distance d (in number o pixels).

5 Palm uses dierent spatial city-block distances d to compute co-occurrence matrices or the classiication o the BarkTex database images : d =, 5,, 5, 2 [3], [5]. e obtains the best classiication results or the distances d =and d =5. That is why we choose to consider D =5dierent distances (d =, 2, 3, 4, 5) to characterize the color texture images o the BarkTex database. The co-occurrence matrices characterize the textures, but they cannot be easily exploited or color texture classiication because they contain a large amount o inormation. To reduce it, while preserving the relevance o these descriptors, aralick proposes to use = 4 eatures, denoted to 4, extracted rom each matrix [2]. IV. FEATURE SELECTIO A. Candidate color texture eatures For each image I coded in a color space S, we compute one single color LBP image LBP8 S [I]. Then, or each o the D =5neighborhoods corresponding to the ive dierent distances, we extract one co-occurrence matrix rom this LBP image (m [ LBP S 8 [I] ] ). Finally 4 aralick eatures are extracted rom each matrix. The number o color spaces used here being equal to S =28, we examine = D S = = 96 color texture eatures denoted x, =,...,. Figure 6 shows how these candidate color texture eatures are extracted. Image I (R, G, B) (L, U, V ) LBP RGB 8 [I] LBP LUV 8 [I] Other color spaces... Fig. 6. m d= [LBP RGB 8 [I]] m d=2 [LBP RGB 8 [I]] m d=3 [LBP RGB 8 [I]] m d=4 [LBP RGB 8 [I]] m d=5 [LBP RGB 8 [I]] m d= [LBP LUV 8 [I]] m d=2 [LBP LUV 8 [I]] m d=3 [LBP LUV 8 [I]] m d=4 [LBP LUV 8 [I]] m d=5 [LBP LUV 8 [I]] Candidate color texture eatures Since the total number o color texture eatures is very high, it is interesting to select the most discriminating ones in order to reduce the size o the eature space and decrease the classiication time. B. Iterative selection The determination o the most discriminating eature space is achieved thanks to an iterative selection procedure based on a supervised learning scheme. This non-exhaustive procedure has given very good results to select an hybrid color space or color image segmentation [4], [7]. In a irst time, ω learning images which are representative o each o the T texture classes are interactively selected by the user. Then, the procedure selects automatically the eatures which discriminate the T texture classes among the = 96 color texture eatures, thanks to the ollowing iterative selection procedure. At each step s o this procedure, an inormational criterion J s is calculated in order to measure the discriminating power o each candidate eature space. At the beginning o this procedure (s =), the one-dimensional candidate eature spaces, deined by each o the available color texture eatures, are considered. The candidate eature which maximizes J is the best one or discriminating the texture classes. It is selected at the irst step and is associated in the second step o the procedure (s =2) to each o the ( ) remaining candidate color texture eatures in order to constitute ( ) two-dimensional candidate eature spaces. We consider that the two-dimensional space which maximizes J 2 is the best space or discriminating the texture classes... In order to only select color texture eatures which are not correlated, we measure, at each step s 2 o the procedure, the correlation between each o the available color texture eatures and each o the s other color texture eatures constituting the selected s dimensional space. The considered eatures will be selected as candidate ones only i their correlation level with the color texture eatures already selected is lower than a threshold ixed by the user [4]. We assume that the more the clusters associated to the dierent texture classes are well separated and compact in the candidate eature space, the higher the discriminating power o the selected color texture eatures is. That leads us to choose measures o class separability and class compactness as measures o the discriminating power. At each step s o the procedure and or each o the ( s+) s-dimensional candidate eature spaces, we deine, or the i th learning image ω i,j (i =,..., ω ) associated to the texture class T j (j =,..., T ), a color texture eature vector X i,j = [x i,j,..., xs i,j ]T where x s i,j is the sth color texture eature. The measure o compactness o each texture class T j is deined by the within-class dispersion matrix Σ C : T ω Σ C = (X i,j M j )(X i,j M j ) T ω T j= i= where M j =[m j,..., ms j ]T is the mean vector o the s color texture eatures o the class T j and ω the number o images by class. The measure o the class separability is deined by the between-class dispersion matrix Σ S : Σ S = T (M j M)(M j M) T T j=

6 where M =[m,..., m s ] T is the mean vector o the s color texture eatures or all the classes. The most discriminating eature space maximizes the inormation criterion : ) J s =trace ((Σ C +Σ S ) Σ S V. EXPERIMETAL RESULTS In order to show the interest o our method or color texture classiication, experimental results are achieved with the color textures o the BarkTex database [5]. Ater having described this benchmark database and shown some examples o LBP images extracted rom BarkTex textures, the results o selection and classiication will be presented and analyzed. A. BarkTex database Color images o the BarkTex database are equally divided into six tree bark classes (Betula pendula (T ), Fagus silvatica (T 2 ), Picéa abies (T 3 ), Pinus silvestris (T 4 ), Quercus robus (T 5 ), Robinia pseudacacia (T 6 )). Each class regroups 68 images o size yielding a collection o 48 images. To build the learning set, we have extracted ω = 32 learning images ω i,j o each texture class T j. For the classiication, 36 test images or each texture class T j are used. Figure 7 illustrates a subset o learning images on the let and a part o the images used to test our classiication method on the right. These test images are classiied thanks to the k-nearest neighbor classiier. We choose to use the -dimensional most discriminating eature space selected by the selection procedure and a number o neighbors k equal to 7 to classiy the test images, because these parameters give the best rate o well-classiied images. B. Examples o LBP images Figure 8 illustrates six color texture images (coded in the (R, G, B) color space) and their associated LBP images. We can notice that these LBP images emphasize not only the pattern o each image, but also the local color variations. Otherwise, contrary to the original images where the textures mainly contain vertical patterns, the patterns present in the associated LBP images have no privileged direction. So, as the choice o the neighborhood used to compute cooccurrence matrices depends on the analyzed textures [2], and as these matrices are extracted rom the LBP images and not directly rom original image, we choose the 8-neighborhood to compute the co-occurrence matrices in order to take into account all directions. C. Selected texture eature space The supervised learning procedure iteratively selects discriminating color texture eatures. Table shows that, at the irst iteration step (s =), the most discriminating color texture eature which maximises J, is the tenth aralick eature extracted rom the co-occurrence matrix m d=3 [LBP (X,Y,Z) 8 ]. This matrix measures the spatial interactions between the pixels o the LBP image LBP (X,Y,Z) 8 in a 8-neighborhood where the distance d between the analysed pixel and its neighbors is equal to d =3. LBP (X,Y,Z) 8 is extracted rom the BarkTex images coded in the (X, Y, Z) color space and the neighborhood taken into account to compute this LBP image is also a 8-neighborhood. The discriminating power o this color texture eature is equal to.758. At the second iteration step, this eature is associated to the aralick eature extracted rom the color co-occurrence matrix m d=2 [LBP (Y,U,V ) 8 ] to constitute the most discriminating two-dimensional eature space with respect to J 2 :the discriminating power o this eature space is equal to Table. Color texture eatures iteratively selected. s Co-occurrence matrix extracted rom the LBP image m d=3 [LBP (X,Y,Z) 8 ] m d=2 [LBP (Y,U,V ) 8 ] m d=4 [LBP (L,Suv,huv) 8 ] m d=2 [LBP (x,y,z) 8 ] m d= [LBP (L,U,V ) 8 ] m d=5 [LBP (A,C,C2) 8 ] m d=4 [LBP (A,C,C2) 8 ] m d= [LBP (x,y,z) 8 ] m d=5 [LBP (r,g,b) 8 ] m d= [LBP (A,C,C2) 8 ] aralick eature J s We stop the iterative procedure at s =and consider the -dimensional most discriminating eature space constitued by the irst ten selected eatures to classiy test images, since this is with this dimension that we obtain the best classiication result. D. Classiication results The rate o well-classiied images obtained by considering the -dimensional eature space above determined reaches 85.6% by classiying test images with a k = 7 nearest neighbor classiier. Since the textures present in the BarkTex database are quite diicult to be discriminated, our method o color texture classiication provides very encouraging results. Indeed, the best classiication result obtained with this benchmark database is 87%, with a 5-dimensional eature space, composed o eatures extracted rom sum and dierence histograms, and a leaving-one-out classiication scheme [22].

7 Learning images Test images T T 2 T 3 T 4 T 5 T 6 Fig. 7. Examples o BarkTex images (a) T class image (b) LBP image computed rom image (a) (c) T 2 class image (d) LBP image computed rom image (c) (e) T 3 class image () LBP image computed rom image (e) (g) T 4 class image (h) LBP image computed rom image (g) (i) T 5 class image (j) LBP image computed rom image (i) (k) T 6 class image (l) LBP image computed rom image (k) Fig. 8. Color texture images o the BarkTex database and their associated LBP images.

8 In order to show the interest to associate dierent color spaces, we have compared the previous rate with the result obtained by considering images only coded in the (R, G, B) space. The best classiication result obtained with the aralick eatures extracted rom co-occurrences matrices computed rom LBP images coded in the single (R, G, B) space reaches 72.7%. This rate is obtained by considering the 8-dimensional most discriminating eature space selected by the iterative selection procedure, and the k =7-nearest neighbor classiier. This experiment conirms that the association o several color spaces improves the characterization o color textures and consequently the results o texture classiication. VI. COCLUSIO The originality o this work lies in the use o a new descriptor to characterize color textures, the color LBP images, which diers rom the color LBP initially proposed by Mäenpää and Pietikäinen. Indeed, instead o comparing the color components o pixels, we compare their color rank thanks to a partial color order relation based on euclidean distances. This approach allows to consider only one LBP by image instead o nine. Another originality is to use aralick eatures extracted rom co-occurrence matrices computed rom the color LBP image to test the eectiveness o this descriptor or color texture classiication. Finally, it is more interesting to extract this color LBP image rom color texture image coded in 28 dierent color spaces. An iterative selection procedure allows then to select among the extracted eatures, those which discriminate the textures, in order to build a low dimensional eature space. Experimental results, achieved with BarkTex database, show the interest o this method with which a satisying rate o well-classiied images (85.6%) is obtained, by analysing a - dimensional eature space. The perspectives o this work are irstly to ind an eicient stopping criterion or the iterative selection procedure, then to determine the number k o neighbors used to classiy test images and inally, to evaluate the relevance o the color order relation. Currently, we apply our approach to control the quality o decorated glasses which can present deects on color texture areas. ACKOWLEDGEMETS This research is unded by Pôle de Compétitivité Maud and Région ord-pas de Calais. REFERECES [] X. Xie. A review o recent advances in surace deect detection using texture analysis techniques. Computer Vision and Image Analysis, 7(3):- 22, 28. [2] T. Mäenpää and M. Pietikäinen. Classiication with color and texture : jointly or separately?. Pattern Recognition, 37 :629-64, 24. [3] C. Palm. Color texture classiication by integrative co-occurrence matrices. Pattern Recognition, 37 : , 24. [4] A. Drimbarean and P.-F. Whelan. Experiments in colour texture analysis. Pattern Recognition, 22 :6-67, 2. [5] S. Chindaro, K. Sirlantzis and F. Deravi. Texture classiication system using colour space usion. Electronics Letters, 4 :589-59, 25. [6] O.J. ernandez, J. Cook, M. Griin, C. De Rama and M. McGovern. Classiication o color textures with random ield models and neural networks. Journal o Computer Science & Technology, 5 :5-57, 25. [7] S. Arivazhagan, L. Ganesan and V. Angayarkanni. Color texture classiication using wavelet transorm. In Proceedings o the Sixth International Conerence on Computational Intelligence and Multimedia Applications, 35-32, 25. [8] A. Sengur. Wavelet transorm and adaptive neuro-uzzy inerence system or color texture classiication. Expert Systems with Applications, doi :.6/j.eswa , 27. [9] G. Van de Wouwer, P. Scheunders, S. Livens and D. Van Dyck. Wavelet correlation signatures or color texture characterization. Pattern Recognition Letters, 32 :443-45, 999. [] T. Ojala, M. Pietikäinen and D. arwood. A comparative study o texture measures with classiication based on eature distributions. Pattern Recognition Letters, 29() :5-59, 996. [] M. Pietikäinen, T. Mäenpää and J. Viertola Color texture classiication with color histograms and local binary patterns In Proceedings o the 2nd International Workshop on Texture Analysis and Synthesis, 9-2, 22. [2] L. Busin,. Vandenbroucke and L. Macaire. Color spaces and image segmentation. Advances in Imaging and Electron Physics, Elsevier, 5(2) :65-68, 28. [3] A. Porebski,. Vandenbroucke and L. Macaire. Iterative eature selection or color texture classiication. In Proceeding o the 4th IEEE International Conerence on Image Processing (ICIP 7), San Antonio, Texas, USA, 3 :59-52, 27. [4]. Vandenbroucke, L. Macaire and J.-G. Postaire. Color image segmentation by pixel classiication in an adapted hybrid color space. Application to soccer image analysis. Computer Vision and Image Understanding, 9 :9-26, 23. [5] R. Picard, C. Graczyk, S. Mann, J. Wachman, L. Picard and L. Campbell. BarkTex benchmark database o color textured images. Media Laboratory, Massachusetts Institute o Technology (MIT), Cambridge, tp ://tphost.uni-koblenz.de/outgoing/vision/lakmann/barktex. [6]. Vandenbroucke, L. Macaire and J.-G. Postaire. Color image segmentation by supervised pixel classiication in a color texture eature space. Application to soccer image segmentation. In Proceeding o the International Conerence on Pattern Recognition (ICPR ), Barcelona, 3 : , 2. [7]. Vandenbroucke, L. Macaire and J.-G. Postaire. Color pixels classiication in an hybrid color space. In Proceeding o the IEEE International Conerence on Image Processing (ICIP 98), Chicago, :76-8, 998. [8] M. Bartkowiak and M. Domanski. Vector median ilters or processing o color images in various color spaces. In Proceeding o the 5th International Conerence on Image Processing and its Applications, , 995. [9] T. Mäenpää, J. Viertola and M. Pietikäinen. Optimising colour and texture eatures or real-time visual inspection. Pattern Analysis and Applications, 6(3) :69-75, 23. [2] R. aralick, K. Shanmugan and I. Dinstein. Textural eatures or image classiication. IEEE Transactions on Systems, Man and Cybernetics, 3 :6-62, 973. [2] A. Porebski,. Vandenbroucke and L. Macaire. eighborhood and aralick eature extraction or color texture analysis. In Proceeding o the 4th European Conerence on Colour in Graphics, Image and Vision (CGIV 8), Terrassa, Spain, 36-32, 28. [22] C. Münzenmayer,. Volk, C. Küblbeck, K. Spinnler and T. Wittenberg. Multispectral texture analysis using interplane sum- and dierencehistograms. DAGM-Symposium, Springer, 42-49, 22.