Second order model deviations of local Gabor features for texture classification

In this paper, we tackle the problem of texture classification with a local approach based on measuring second order deviations with respect to a dictionary of characteristic patterns. At each pixel, we extract local signal properties thanks to several Gabor filters that are aggregated on a small support region. Then, we compute a dictionary of such features that serves as a universal model. The texture signature is the deviation of second order statistics between its local features and the universal model. Experiments are made on two sets of photographic paper textures, and show the soundness of the approach.