Constraint score for semi-supervised selection of color texture features M. Kalakech *† , A. Porebski † , P. Biela *† , D. Hamad ‡ and L. Macaire † * Hautes Etudes d’Ingénieur - 13 rue de Toul - 59046 Lille - FRANCE e-mail: mariam.kalakech@hei.fr, philippe.biela@hei.fr † Laboratoire LAGIS - FRE CNRS 3303 - Université Lille 1 - 59655 Villeneuve d’Ascq - FRANCE e-mail: alice.porebski@univ-lille1.fr, ludovic.macaire@univ-lille1.fr ‡ Laboratoire LISIC - Université du Littoral Côte d’Opale - 62228 Calais - FRANCE e-mail: denis.hamad@lisic.univ-littoral.fr Abstract—In this paper, we propose a semi-supervised color texture feature selection scheme based on a new score. This score uses both unlabeled textures and pairwise constraints which specify whether a pair of textures belongs or not to the same class. Experimental results with benchmark databases show that introducing a few constraints improves the performances reached by a color texture classification scheme. Index Terms—feature selection; Haralick features; color tex- ture; pairwise constraints. I. I NTRODUCTION Many authors have shown that the use of color improves the results of texture classification compared with grey level image analysis. Moreover, there exists a large number of color spaces which respect different properties and it is well known that the choice of the color space impacts the quality of texture classification. The prior determination of a color space which is well suited to the considered texture discrimination, is not easy. So, we follow a multi color space approach where the properties of several different color spaces are together considered in order to analyze color textures. Van Den Broek et al. have shown that there does not exist any color texture features which allow to discriminate all kinds of color textures [11]. That is why we propose a scheme which selects texture features among a high number of available color texture features from images coded in several color spaces. The key point of feature selection is the definition of score which evaluates the relevance of features, with respect to the considered texture samples. Real industrial applications of texture classification mainly concern aspect control. They have huge texture samples and the label assignment by a user is tedious. In that case, one performs unsupervised feature selection without any supervi- sion information about the labels of the texture samples. The feature selection scheme is based on unsupervised scores like Variance or Laplacian scores [7]. Though, it would be interesting to use partial and incomplete prior knowledge about the texture samples in order to improve the quality of texture selection. For this purpose, we develop a semi-supervised feature selection scheme which analyzes huge texture samples without any prior knowledge and a small supervision information provided by the user feedback. Beside the classical user supervision information described by class labels, the user can define the pairwise constraints between a few texture samples. It consists to simply specify whether a pair of textures samples must be regrouped to- gether (must-link constraints) or cannot be regrouped together (cannot-link constraints), without needing to identify the dif- ferent texture classes. Our semi-supervised feature selection scheme requires to evaluate the relevance of each texture feature thanks to a score which uses both unlabeled textures samples and pairwise constraints. In this paper, we measure the accuracy improvement brought by a few constraints to select features. For this purpose, we compare the quality of tex- ture classification obtained by an unsupervised classification scheme and by our proposed semi-supervised classification scheme. The paper is organized as follows. Used Color texture features are introduced in section II. In section III, we review the unsupervised variance and Laplacian scores and introduce our semi-supervised constraint score. Experimental results, based on three well-know benchmark texture databases, are finally presented in section IV. II. COLOR TEXTURE FEATURES The color of pixels can be represented in different color spaces which respect different physical, physiologic, and psycho-visual properties [1]. However, there is no color space which is adapted to the discrimination of all textures. To solve this problem, we propose to combine the acquisition (R, G, B), perceptual (H, S, V ) and well known (L * ,a * ,b * ) color spaces in order to improve the performances of color image classification. We associate the informations coming from different color spaces by coding color texture images into these color spaces and by computing some texture features from the so-coded images. These color spaces are retained since the RGB transformation is a non linear one. Porebski et al. [10] have shown that the well known Haralick features extracted from Reduced Size Chromatic Co- occurrence Matrices (RSCCM) are relevant to discriminate color texture classes. That is why, we choose these features 2010 The 3rd International Conference on Machine Vision (ICMV 2010) 275 ISBN: C 978-1-4244-8889-6 ICMV 2011