Texture Classification via Local Feature Representation of Multi-order Gradients

This paper presents a novel method to texture classification using local feature representation of multiple order gradients. Different from the state of the art approaches in literature that make use of the widely-used first order gradient based local descriptors, e.g. LBP, HOG, DAISY, SIFT, etc., we claim that the second order gradient based ones also provide critical contribution to classification performance, and thus propose to use Histogram of Second Order Gradients (HSOG) to describe micro-texton patterns. Both the similarity measurements of first and second order gradients computed by Bag-of-Feature modeling and SVM classifier are combined for decision making. Experimental results achieved on the Outex_TC dataset not only illustrate that the second order gradient based HSOG is effective to classify texture images, but also highlight that multiple order gradient based description by fusing complementary clues of the first and second order gradients is a promising solution to improve the accuracy in texture classification.