FRAME BASED TEXTURE CLASSIFICATION BY CONSIDERING VARIOUS SPATIAL NEIGHBORHOODS. Karl Skretting and John Håkon Husøy

Transcription

1 FRAME BASED TEXTURE CLASSIFICATION BY CONSIDERING VARIOUS SPATIAL NEIGHBORHOODS Kar Skretting and John Håkon Husøy University of Stavanger, Department of Eectrica and Computer Engineering N-4036 Stavanger, Norway, E-mai: ABSTRACT In this paper we briefy describe the Frame Texture Cassification Method (FTCM) and the tied-foor texture mode on which it is based. For a test image pixe a feature vector is formed based on a sma spatia neighborhood, and FTCM cacuate its distance to the frames, i.e. the dictionaries, one frame for each texture cass under consideration. The frame where the distance is smaest gives the texture cass the pixe is assumed to beong to. The texture mode gives some guidance to how the feature vectors shoud be formed, many different ways to do this are described, a being in accordance with the texture mode. In the experiments we investigate many different sizes and shapes for the neighborhood used when forming the feature vectors. The experiments are done on some commony used test images of natura textures. FTCM is fexibe when seecting a neighborhood and many patterns give exceent cassification resuts. 1. INTRODUCTION Cassification and segmentation of texture content in digita images is an important part of many image anaysis methods. Tuceryan and Jain [1] give a comprehensive overview of texture cassification. Possibe appications can be grouped into: 1) texture anaysis, i.e. find some appropriate properties for a texture, 2) texture cassification, i.e. identify the texture cass in a homogeneous region, 3) texture segmentation, i.e. find a boundary map between different texture regions of an image. The boundary map may be used for object recognition and scene interpretation in areas such as medica diagnostics, geophysica interpretation, industria automation and image indexing. Finay, 4) texture synthesis, i.e. generate artificia textures to be used for exampe in computer graphics or image compression. Some exampes of appications are presented in [2, 3, 4, 5, 6]. For human beings it is quite easy to recognize different textures in images. We ook at the whoe image at once and find the texture features at the same time as we do image segmentation. Both these tasks are highy knowedge dependent, we know a ot of common textures and know in which shapes and contexts they usuay occur. Computer based agorithms do not (yet) have this arge amount of knowedge avaiabe and typicay do texture cassification and image segmentation in two ceary separated parts. Aso, texture cassification agorithms typicay have two main parts: A oca feature vector is found, which is subsequenty used for texture cassification or segmentation. The feature vector is generated, a ot of different methods can be used, from the pixes in a spatia neighborhood of a center pixe. The methods for feature extraction may be oosey grouped as statistica, geometrica, mode-based, and signa processing (fitering) methods [1]. For the fitering methods the feature vectors are often buit as variance estimates, oca energy measures, for each of the sub-bands of a fiter bank. Aso, there are numerous cassification or pattern recognition methods avaiabe: The Bayes cassifier is probaby the most common one [7, 8]. The minor max-seector is a simpe one that can be used if each entry in the feature vector measures the simiarity to, or corresponds to, a texture cass. Nearest neighbor cassifier, vector quantization (codebook vectors represent each cass) [9] and earning vector quantization (LVQ) (codebook vectors define the decision borders) [10, 11, 12], neura networks, watershed-based agorithm [13], and support vector machines (SVM) [14] are other methods. In this paper Frame Texture Cassification Method (FTCM) shoud be regarded as a supervised vector cassification method. As wi be expained in Section 2 FTCM is particuary we suited under the assumption that the vectors within each cass are a sparse inear combination of eements from a frame, i.e a dictionary, of finite (moderate) size. This assumption wi be true for feature vectors generated from a texture synthesized by the tied-foor texture mode. FTCM is thus reated to other texture cassification methods which depend on cassification of the feature vectors. It is possibe that the resuts presented in this paper may be usefu aso when other vector cassification methods are used for texture cassification.

2 We wi particuary mention the reationship between FTCM and the SVM scheme as used in [15]. SVM finds a set of support vectors for each texture and this set identifies a hyperpane which separates the given texture from the rest of the textures, whie FTCM finds a set of frame vectors for each texture and this set is trained to efficienty represent the given texture by a sparse inear combination and thus identifying the texture. 2. FRAME TEXTURE CLASSIFICATION METHOD The FTCM is a supervised texture cassification method. It was first introduced in [16], and is aso presented in [17] and [18]. Here foows a brief presentation of the underaying deterministic texture mode, foowed by the two parts of FTCM The tied-foor texture mode The main resut of this subsection is that it is reasonabe to mode a sma texture image bock as a sparse inear combination of frame eements. In the proposed texture mode a texture is modeed as a tied foor, where a ties are identica. We et the coordinate system be aigned to match a tie, such that the center of the first tie is given by (x, y) = ( 1 2, 1 2 ), and the corners are (0, 0), (1, 0), (0, 1), and (1, 1). The coor, or gray-eve, at a given position on the foor is given by a continuous periodic two-dimensiona function which we denote c(x, y) = c(x x, y y ). A finite number, denoted M, of contro points, denoted c i, are paced on each tie. The function is defined as a biinear interpoation of four of the (cosest) contro points, i.e. c(x, y) = a 1 c i1 + a 2 c i2 + a 3 c i3 + a 4 c i4. The biinear interpoation is actuay a convex combination, with a 1 +a 2 +a 3 +a 4 = 1 and 0 a k 1. Sampes of c(x, y) on a rectanguar samping grid, not necessariy aigned with the coordinate system impied by the first tie, constitute the digita texture image. By choosing 1) the number and positions of contro points in a tie, 2) the gray-eve vaue (coor) of each of the contro points, 3) the orientation of the samping grid reative to the coordinate system aigned with the ties, denoted by ange α, and finay, 4) the distance between neighboring samping points, denoted by δ, in the samping grid, we obtain a digita texture image. We wi now ook coser on a sma bock of pixes from the texture image. This bock is arranged into a size N vector, x = [x(1), x(2),..., x(n)] T, x(1) is the upper eft pixe and the rest are numbered coumnwise. Remember that each entry of x, i.e. each pixe, is a convex combination of four contro points taken from a finite set of M contro points. Given the mode above, it was shown in [18] that the vector x is a convex combination of four frame vectors taken from a finite set of at most MN 2 frame vectors Frame design For a particuar texture image, specified by the mode above, i.e. by contro points and by a samping grid, we coud find the correct frame. But in FTCM a more genera method is used to design the frame, the frame is not designed based on knowedge on how the texture image was made, but ony based on a set of training vectors generated from an exampe image. Each training vector is approximated by a inear combination of (three or) four frame vectors. The frame is designed to achieve the best tota sparse representation for a the training vectors. How the training vectors are formed is discussed in Section 3. The parameters used during design are N, K, L and s. The ength of the training (and frame) vectors is N. The number of vectors in the frame is denoted K. For practica reasons it shoud be quite sma, we may use N K 10N MN 2. To avoid overtraining the number of training vectors L shoud be much arger than K, K L. The sparseness to use in the approximation is denoted s, here we use the vaues s = 3 and s = 4. For each texture cass the training vectors are arranged into a N L matrix, X = [x 1 x 2... x L ]. We want to find a frame with K frame vectors, F = [f 1 f 2... f K ], such that each of the training vectors can be approximated by a sparse combination of the frame vectors, x Fw where w has at most s non-zero entries. Having W = [w 1 w 2... w L ] the objective function to be minimized, with a sparseness constraint on W, is J = J(F, W) = X FW 2. (1) Finding the optima soution to this probem is difficut if not impossibe. We spit the probem into two parts to make it more tractabe, simiar to what is done in the GLA design agorithm for VQ codebooks [19]. The iterative soution strategy presented beow resuts in good, but in genera suboptima, soutions to the probem. The agorithm starts with a user suppied initia frame F 0, usuay K arbitrary vectors from the set of training vectors, and then improves it by iterativey repeating two main steps: 1. W t is found by vector seection using frame F t. The objective function is J(W) = X F t W 2, and a sparseness constraint is imposed on W. 2. F t+1 is found from X and W t, where the objective function is J(F) = X FW t 2. This

3 gives: F t+1 = XWt T ( Wt Wt T ) 1 (2) Then we increment t and go to step 1. t is the iteration number. The first step is suboptima due to the use of practica vector seection agorithms, whie the second step finds the F that minimizes the objective function Cassification Texture cassification of a test image, containing regions of different textures, is the task of cassifying each pixe of the test image to beong to a certain texture. This is done by generating test vectors from the test image using the same method as when the training vectors were generated. A test vector is represented in a sparse way using each of the different frames that were trained for the textures under consideration, the set of C frames {F (i) }. The distance between test vector x and frame F (i) is the 2-norm (ength) of the representation error for the best sparse representation of the test vector, d (i) = r (i) and r (i) = x F (i) w (i). Each test vector x corresponds to a pixe of the test image. Cassification consists of seecting the index i for which the norm, or norm squared, of the representation error, r (i) 2 = r (i)t r (i), is minimized. Direct cassification based on the norm squared of the representation error for each test vector (pixe) gives quite arge cassification errors, but the resuts can be substantiay improved by smoothing the error images. Smoothing is reasonabe since it is ikey that neighboring pixes beong to the same texture. For smoothing we have used the separabe Gaussian ow-pass fiter. The effect of smoothing is mainy that more smoothing gives better cassification within the texture regions but ower resoution, i.e. often more cassification errors aong the borders between different texture regions. To improve texture segmentation a noninearity may be incuded before the smoothing fiter is appied, [20]. We have found that the ogarithmic noninearity often gives the best resuts, [16]. 3. SELECTING PIXELS IN LOCAL NEIGHBORHOOD We wi here show that the training and test vectors (feature vectors) can be made in many different ways and sti be in accordance with the mode in subsection 2.1. Let us assume that the bock around each pixe is of size 13 13, giving N = 169 as the ength of vector x. The perfect frame is F of size N K where K = MN 2, M is the number of contro points on one tie. In accordance with the mode we have x = Fw, where w is a sparse vector with ony four non-zero entries. Let another feature vector x be formed as x = Ax where A is a N N matrix, N N. Keeping w the same as above, i.e. sti sparse, this gives x = A(Fw) = (AF)w = F w, and we see that aso x can be represented by a inear combination of four frame vectors taken from a finite set of frame vectors, F = AF. Suppose each row in A has ony one non-zero entry. This gives the case where the feature vectors are formed by a pattern of pixes around the center pixe. Some exampes are iustrated in patterns 1-9 in Figure 1. The non-zeros entries of A are here assumed to be 1, but it is aso possibe to form a weighted pattern. Without restrictions on the entries of A each eement of x is a inear combination of the eements in x. This is the same as the output of different FIR-fiters appied on the pixes in the oca neighborhood of the centra pixe, each row of A gives the coefficients of each fiter. The patterns in Figure 2 are formed this way. In the mode a vector is formed by a biinear interpoation of frame vectors, i.e. a sparse convex combination where each coefficient 0 a k 1 and ak = 1. In vector seection, and the equation x = Fw, a sparse inear combination is used. The coefficients, a k, are the non-zero entries of w. It is not trivia to incude the convex constraint into the vector seection agorithm, but an affine constraint, i.e. ak = 1, can be impemented without any changes to the inear vector seection agorithm. Let x = [ b x ] [ and F b b b = f 1 f 2 f K ]. (3) where b 0. Then the inear combination x = F w must aso be an affine combination. Patterns 10-13, 15, 16 and 18 in Figure 2 a have an extra fixed eement added. Pattern 10 is ike pattern 1 except for the fixed eement, 11 is ike 5, 12 is ike 3, 13 is ike 6, 15 is ike 14, and 18 is ike 17. Here we have argued that the feature vectors may be generated in many different ways from the pixes in a spatia neighborhood of the centra pixe. Neither the mode nor the theory give any cear guidance to which is better. Obviousy each vector shoud contain the necessary texture information for this particuar center pixe, but it does not need to contain information on the neighboring pixes (smoothing). For the compexity of the methods, the smaer feature vectors shoud be preferred. In the next section extensive experiments are presented, a the 18 different methods shown in Figure 1 and Figure 2 are tested.

4 Pattern 1, N=9 Pattern 2, N=16 Pattern 3, N=25 (a) (b) Pattern 4, N=49 Pattern 5, N=13 Pattern 6, N=21 Figure 3: Two of the nine test images used. 4. EXPERIMENTS Pattern 7, N=21 Pattern 8, N=49 Pattern 9, N=39 Figure 1: Pixes around the center pixe, marked by a square, used to form the oca feature vector for the 9 simpe cases. N is the number entries in the feature vector, here it the same as the number of pixes in each pattern. Pattern 10, N=10 Pattern 11, N=14 Pattern 12, N=26 Pattern 13, N=22 Pattern 14, N=25 Pattern 15, N=26 Pattern 16, N=25 Pattern 17, N=17 Pattern 18, N=18 Figure 2: More patterns, here an extra fixed vaue eement is added to each feature vector, except for pattern 14 and 17. Feature vectors for patterns 14 to 18 are formed by fitering over the pixes indicated by pus sign, for pattern 17 and 18 the Daubechies 7-9 waveet fiters are used. N is the number entries in the feature vector. FTCM and the 18 different patterns for the feature vectors, described in Section 3, were tested both for synthesized texture images made based on the mode and for natura rea textures. It turned out that a patterns achieved very we on the synthesized test images, and the differences between them were margina. Thus, this section presents the resuts for the natura textures. As in [18] we choose to use the nine test images of Randen and Husøy [11] for natura rea textures. These consist of 77 different natura textures, taken from three different and commony used texture sources: The Brodatz abum, the MIT Vision Texture database, and the MeasTex Image Texture Database. The test images are denoted (a) to (i). (a) and (b) are shown in Figure 3, a are shown in Figure 11 in [11], where aso a more detaied description of the test images can be found 1. The same test images were aso used in other papers [21, 8, 13, 15, 22]. The exceent performance of FTCM compared to the other methods was reported in [18], the purpose here is to compare the different patterns to each other. The experiments started by designing the frames to be used, one frame for each kind of training vector and texture under consideration. The frame parameters, see Subsection 2.2, used were: N as given by the pattern, K = 6N but minimum 100 and maximum 200, s = 3 for N < 30 and s = 4 for N 30, and L = iterations were done. During cassification of each test image we use ony the frames corresponding to the textures that are present in the test image. A ogarithmic noninearity and a separabe Gaussian ow-pass fiter were used. The standard deviation σ in the fiter varied, but here we present the compete resuts ony for σ = 10, the size of this ow pass fiter was cut at approximatey 2σ at each end, giving the fiter size pixes. These resuts are shown in Tabe 1. 1 The training images and the test images are avaiabe at tranden/

5 Pattern (parameters: N K, s) a b c d e f g h i Mean 1 (9 100, 3) (16 100, 3) (25 150, 3) (49 200, 4) (13 100, 3) (21 126, 3) (21 126, 3) (49 200, 4) (39 200, 4) (10 100, 3) (14 100, 3) (26 156, 3) (22 132, 3) (25 150, 3) (26 156, 3) (25 150, 3) (17 102, 3) (18 108, 3) , see text Tabe 1: Cassification errors, given as percentage wrongy cassified pixes, for different methods and natura test images. Let us ook coser at the average resuts for the 9 test images. These are presented in Figure 4, on the eft side for σ = 1, here ow-pass fitering is done using a sma 3 3 and narrow fiter, and on the right side for σ = 10. We note that with itte ow-pass fitering the onger feature vectors perform best, whie using more appropriate ow-pass fitering it seems that onger feature vectors do not improve the resuts. The most probabe expanation is that the onger feature vectors aso contain some of the smoothing information. We aso concude that amost a methods seems to capture the essentia texture information quite we. Methods 11, 15, and 18 do a very we. To see how an extra constant eement added to the feature vector infuence on the resuts we compare the resuts, taken from Tabe 1, of the six patterns, 1, 5, 3, 6, 14 and 17 to the six simiar patterns, 10, 11, 12, 13, 15 and 18. We note that adding an extra eement give better resuts amost aways, the means (of the means) are 14.6 and 12.7 percent wrongy cassified pixes. The concusion seems to be cear, adding the extra eement improves the resuts. This is in ine with the mode where an affine combination is preferred to a inear combination. Looking at the detais the concusion is not that cear, three of the test images, c, e and g, are best cassified using method 14. Generay, the detais give more scattered resuts, indicating that the choice of pattern is dependent on the test image. Aso, training has a random eement both in seection of the training vectors, the initia frame, and the training process, resuting in sighty different frames each time a frame is trained. This random effect obviousy has an impact on the resuts, and can expain some of the scattering. More work is needed to understand this better. To get even better resuts it is possibe to do averaging without smoothing. This is possibe since severa representation errors for each pixe, one for each of the patterns, are avaiabe. Averaging these for patterns 11, 15, and 18 before smoothing (and before the ogarithmic noninearity is appied) gives the resuts presented in ast ine of Tabe CONCLUSION Many of the different sizes and shapes for the neighborhood used to form the feature vectors gave exceent resuts when FTCM was used for vector cassification. It seems that the sma 3 3 pattern, used in method 1 and 10, is too sma. Adding just a few more pixes, as in pattern 2, 5 and 11, seems to capture the texture information we and gives very good resuts. Pattern 11 seems best, considering both simpicity and resuts. To get optima cassification resuts severa patterns can be combined. 6. REFERENCES [1] M. Tuceryan and A. K. Jain, Texture anaysis, in Handbook of Pattern Recognition and Computer Vision, C. H. Chen, L. F. Pau, and P. S. P. Wang, Eds. Singapore: Word Scientific Pubishing Co, 1998, ch. 2.1, pp

6 65 1 Mean for the 9 test images, σ = Mean for the 9 test images, σ = Percentage wrongy cassified pixes Percentage wrongy cassified pixes Length of feature vectors, N Length of feature vectors, N. Figure 4: The average percentage wrongy cassified pixes for the 9 test image, here with σ = 1 in ow-pass fiter to the eft and σ = 10 to the right. [2] R. J. Dekker, Texture anaysis and cassification of ERS SAR images for map updating of urban areas in the Netherands, IEEE Trans. Geosci. Remote Sensing, vo. 41, no. 9, pp , Sept [3] M. K. Kundu and M. Acharyya, M-band waveets: Appication to texture segmentation for rea ife image anaysis, Internationa Journa of Waveets, Mutiresoution and Information Processing, vo. 1, no. 1, pp , [4] F. Mendoza and J. M. Aguiera, Appication of image anaysis for cassification of ripening bananas, Journa of Food Science, vo. 69, no. 9, pp , Nov [5] S. Arivazhagan and L. Ganesan, Automatic target detection using waveet transform, EURASIP Journa on Appied Signa Processing, vo. 2004, no. 17, pp , [6] S. Singh and M. Singh, A dynamic cassifier seection and combination approach to image region abeing, Signa Processing Image Communication, vo. 30, no. 3, pp , Mar [7] M. Unser, Texture cassification and segmentation using waveet frames, IEEE Trans. Image Processing, vo. 4, no. 11, pp , Nov [8] S. Liapis, E. Sifakis, and G. Tziritas, Coour and texture segmentation using waveet frame anaysis, deterministic reaxation, and fast marching agorithms, Journa of Visua Communication and Image Representation, vo. 15, no. 1, pp. 1 26, Mar [9] G. F. McLean, Vector quantization for texture cassification, IEEE Trans. Syst., Man, Cybern., vo. 23, no. 3, pp , May/June [10] T. Kohonen, The sef-organizing map, Proc. IEEE, vo. 78, no. 9, pp , Sept [11] T. Randen and J. H. Husøy, Fitering for texture cassification: A comparative study, IEEE Trans. Pattern Ana. Machine Inte., vo. 21, no. 4, pp , Apri [12] C. Diamantini and A. Spavieri, Quantizing for minimum average miscassification risk, IEEE Trans. Neura Networks, vo. 9, no. 1, pp , Jan [13] N. Mapica, J. E. Ortuño, and A. Santos, A mutichanne watershed-based agorithm for supervised texture segmentation, Pattern Recognition Letters, vo. 24, no. 9-10, pp , June [14] S. Li, J. T. Kwok, H. Zhu, and Y. Wang, Texture cassification using the support vector machines, Pattern Recognition, vo. 36, no. 12, pp , Dec [15] K. I. Kim, K. Jung, S. H. Park, and H. J. Kim, Support vector machines for texture cassification, IEEE Trans. Pattern Ana. Machine Inte., vo. 24, no. 11, pp , Nov [16] K. Skretting, Sparse signa representation using overapping frames, Ph.D. dissertation, NTNU Trondheim and Høgskoen i Stavanger, Oct. 2002, avaiabe at karsk/. [17] K. Skretting and J. H. Husøy, Texture cassification using sparse frame based representations, in NORSIG-02, Tromsø/Trondheim, Norway, Oct. 2002, aso avaiabe at karsk/. [18], Texture cassification using sparse frame based representations, EURASIP Journa on Appied Signa Processing, 2005, accepted for pubiction, to be printed autumn [19] A. Gersho and R. M. Gray, Vector Quantization and Signa Compression. Norwe, Mass., USA: Kuwer Academic Pubishers, [20] M. Unser and M. Eden, Noninear operators for improving texture segmentation based on features extracted by spatia fitering, IEEE Trans. Syst., Man, Cybern., vo. 20, pp , [21] M. Acharyya, R. K. De, and M. K. Kundu, Extraction of features using M-band waveet packet frame and their neuro-fuzzy evauation for mutitexture segmentation, IEEE Trans. Pattern Ana. Machine Inte., vo. 44, no. 12, pp , Dec [22] T. Ojaa, K. Vakeaahti, E. Oja, and M. Pietikäinen, Texture discrimination with mutidimensiona distributions of signed gray-eve differences, Pattern Recognition, vo. 34, no. 3, pp , Mar