Self-Similarity Based Classification of 3D Surface Textures

This paper presents a novel 3D surface texture classification method based on self-similarity maps which are calculated directly from raw captured texture images. 3D surface textures have special properties for they are sensitive to illumination and view conditions. Some previous classification methods which are illumination invariant or rotation invariant have shown to be effective to this particularity, but most of them are model-based. We introduce a new approach to classify 3D surface texture image sets by automatically extracted self-similarity maps. Feature vectors are then generated from responses to a filter bank of the self-similarity maps. Classification is achieved by comparing the L2 distance between training and testing feature vectors. The experiment results show our approach achieves an accuracy of 95%.