Color Texture Classification under Different Illuminations Using Rank Correlation Matrices

Color has been shown to be useful in the context of texture classification. However, since under different illuminations color is not stable, color invariant descriptors should be defined when the illumination of the query is unknown. In this paper, we propose to characterize color textures by analyzing the rank correlation of color planes between pixels locally close to each other. Thus, considering one distance and one direction in the image space, we obtain a correlation measure which i) is related to the colors of the pixels, ii) is illumination invariant, iii) represents the spatial interactions between different color components of neighbored pixels. Furthermore, we show how this measure can be very fast extracted from co-occurrence matrices. The discriminative power of this descriptor is assessed on a public color texture database. INTRODUCTION: In this paper, we specifically address the problem of color texture classification across illumination changes. For this purpose, we consider images of color textures acquired with the same viewpoint and the same scale factor but under three different illuminations [33]. In this context, there exist different approaches for color texture For example, the structural approach consists in analyzing the relative positions of features extracted from the image [12]. One other approach tries to model the spatial repartition of the colors in the image. In this aim, one can use Markov Random Fields [30, 27] or Local Binary Pattern [32, 31].