Integrating local feature and global statistics for texture analysis

A main challenge for texture analysis is to construct a compact texture descriptor which is not only highly discriminative to intra-class textures, but also robust to inter-class variations, geometric and photometric changes. In this paper, a new texture descriptor is developed by integrating the local affine-invariant texture features and the global viewpoint-invariant statistics. Based on the pixel clustering using two state-of-art robust local texture descriptors (i.e. SIFT and SPIN), the proposed texture descriptor enables impressive invariance to a wide range of environmental changes (e.g. view changes, illumination changes, surface distortions) by characterizing the spatial distribution of pixel sets using multi-fractal analysis. Experiments on some real datasets (publicly available) showed that the proposed texture descriptor achieved better performance than some state-of-art techniques in texture retrieval and texture classification while the computation cost is significantly reduced.