Texture classification using local discriminative features and Fisher encoding

In this paper we introduce a new image representation for texture classification. Our work is motivated by recent developments in the field of local patch based features, compressive sensing and descriptor encoding methods. Novel features called Compressed Random Pixel Difference (CRPD) are proposed. These features are low in dimensionality, highly discriminative, and easy to compute. Combined with an efficient encoding method, an expressive and robust image descriptor is obtained. Experiments conducted on widely used texture datasets (KTH-TIPS-2a and Brodatz) demonstrate an efficiency of the proposed approach. On KTH-TIPS-2a dataset we have achieved the highest recognition accuracy (to the best of our knowledge) and on Brodatz dataset achieved performance is comparable to the state-of-the-art methods.