Neighborhood and Haralick feature extraction for color texture analysis

Abstract

In this paper, we propose to study the influence of the neighborhood used to process the color co-occurrence matrices on the quality of texture analysis. First, we measure the discriminating power of Haraiick features extracted from the color co-occurrence matrices of color images coded in 28 different color spaces, and we select the most discriminating one for different 3x3 neighborhoods. Then, we experimentally verify that the most discriminating feature space, built by using an iterative selection procedure, depends on the chosen neighborhood and finally we study the impact of the neighborhood choice on the classification results by using the same feature space but different neighborhoods. Experimental results achieved with the Barktek database have firstly shown the adequacy between the discriminating power of the selected feature space and the rate of well-classified images. We have also seen that the choice of the neighborhood does not highly influence the selection of the most discriminating feature but has a significant impact on the quality of discrimination between the considered textures. Indeed, we have worked with textures which contain vertical patterns and have shown that the best classification results have been obtained with horizontal neighborhoods. The choice of the neighborhood depends consequently on the analysed textures. © 2008 Society for Imaging Science and Technology.