358 PRZEGLĄD ELEKTROTECHNICZNY (Electrical Review), ISSN 0033-2097, R. 88 NR 11a/2012 Erkan TANYILDIZI Firat University, Technology Faculty, Department of Software Engineering, 23100, Elazig A hybrid color texture image classification method based on 2D and semi 3D texture features and extreme learning machine Abstract. Color texture classification is an important step in image segmentation and recognition. The color information is especially important in textures of natural scenes. In this paper, we propose a novel approach based on the 2D and semi 3D texture feature coding method (TFCM) for color texture classification. While 2D TFCM features are extracted on gray scale converted color texture images, the semi 3D TFCM features are extracted on RGB coded color texture images. The proposed approach is tested on two publicly available datasets. Moreover, comprehensive comparisons are realized with traditional texture analysis tools. The results show the advantages of the proposed method over other color texture analysis methods. Streszczenie. W artykule zaproponowano nowa metodę klasyfikacji obrazów z kolorowa teksturą wykorzystującą wykorzystującą metody kodowania tekstury 2D. Metodę testowano na dwóch przykładach baz danych i porównano z metodami dotychczas stosowanymi. (Hybrydowa metoda klasyfikacji obrazów z kolorowa teksturą) Key words: Color texture classification, Texture feature coding method, feature extraction Słowa kluczowe: tekstura, klsyfikacja obrazów z teksturą 1. Introduction Texture analysis is very useful for experiments of image classification and identification. Thus, it has long been an area of computer vision with active research area spanning image processing, pattern recognition, and computer vision, with applications to medical image analysis, remote sensing, object recognition, industrial surface inspection, document segmentation and content-based image retrieval. Texture classification has received significant attention with many proposed approaches, as documented in comprehensive surveys [1-5]. The ability of a human to distinguish different textures is apparent, therefore, the automated description and recognition of the texture images is in demand. Over the years, many researchers have studied different texture analysis methods. Many of these methods represent the local behavior of the texture via statistical [6], structural [7] or spectral [8] properties of the image. A methodology is presented in [9], where second-order probability distributions [2, 4] are enough for human discrimination of two texture patterns, has motivated the use of statistical approaches. On the other hand, structural approaches describe the textures by rules, which govern the position of primitive elements, which make up the texture [10]. In addition, signal processing methods, such as Wavelet transform [11-13], Fourier analysis [8] and Gabor filters [14], were motivated by psychophysical researches, which have given evidences that the human brain does a frequency analysis of the image [15, 16]. These approaches represent the texture as an image in a space whose coordinate system has an interpretation that is closely related to the characteristics of a texture. The texture feature coding method (TFCM) that forms the basis for texture features was first discussed by Horng [17] and later applied in various application such as tumor detection and landmine detection [18, 19]. TFCM is a new texture analysis scheme which transforms an original image into a texture feature image whose pixel values represent the texture information of the pixel in original image. The TFCM is a coding scheme that transforms an image into a feature image, in which each pixel is encoded by TFCM into a texture feature number (TFN) that represents a certain type of local texture. The TFN of each pixel in the feature image is generated based on a 3×3 texture unit as well as the gray-level variations of its eight surrounding pixels. The TFN histogram and TFN co-occurrence matrix are derived to generate many texture features for texture classification. The method has several remarkable advantages including accurate representation and record of target texture, and computational efficiency [19]. In this paper, we propose a hybrid method where both 2D and semi 3D TFCM and Extreme Learning Machine (ELM) is combined for efficient color texture classification. ELM was proposed as an alternative and effective approach for neural networks. We firstly review the TFCM and ELM methods and then we conduct several experiments on the various color textures for showing the efficiency of the proposed hybrid scheme. Experimental results are promising and the comparisons show the superiority of our proposal. The rest of this paper is organized as follows. Section 2 reviews the related works for texture classification. Sections 3 and 4 review the TFCM based feature extraction mechanism and the ELM classifier. In Section 5, we evaluate the capabilities of the proposed features with extensive experiments on two texture datasets, and present comparisons with current methods. Finally, we conclude the paper in section 6. 2. Related works Up to now, numerous feature extraction approaches have been proposed for texture image classification. Scale-invariant feature transform (SIFT) [20], Histogram of Oriented Gradients (HOG) feature and the related methods are some of the representative approaches that are widely used in image processing community [21, 22]. Moreover, Chellappa et al. used Gaussian–Markov random field (GMRF) based features to find some statistical relationships among adjacent pixels [23]. Kashyapand Khotanzad then proposed the isotropic circular Gaussian–Markov (ICGMRF) to achieve rotation invariant texture description [24]. Another extension was the isotropic circular GMRF (ACGMRF) designed by Deng and Clausi to encode relative orientations of adjacent pixels [25]. In addition, methods based on multi-channel filtering or wavelet decomposition [26] were also studied. Varma and Zisserman [27] introduced text on histograms in MR8 filtered response space as features. Sengur [11, 12] used the wavelet transform, entropy and energy features for color texture image classification with various classifiers. Local binary pattern (LBP), which is considered as an effective texture classification methodology, was proposed by Ojala et al. [28]. It has many properties such as rotation invariance and low computational cost [29, 30]. Karabatak et al. investigated the usage of association rules on wavelet domain for efficient texture classification [13]. Association rules are robust in modeling the relationship in a given database. Thus, the adjacent pixels interactions in texture structure are modeled in wavelet domain by association rules. A novel approach based on the fractal dimension for color texture analysis was proposed by Backes et al. [31]. The proposed approach investigates the complexity in R, G and B color channels to characterize a texture sample. The authors considered both all channels in combination and the correlations between them. Liu and Fieguth [32] introduced the use of random projections (RPs), a universal, information-preserving dimensionality-reduction technique, to project the patch vector space to a compressed patch space without a loss of salient information, claiming that the performance achieved by random features can outperform patch features, MR8, and LBP features. 3. Texture Feature Coding Method (TFCM) The justification behind the TFCM technique, as proposed by Horng in [17], is the translation of an intensity image to a texture feature number (TFN) image via differencing in the image domain followed by successive stages of vector classification.