A study of regional distributions and dissimilarity measures for multi-scale nonlinear structure tensor in texture segmentation

To represent the orientation and scale differences of texture images effectively, multi-scale nonlinear structure tensor (MSNST) has been recently proposed to extract the texture features in our previous research [1]. In this paper, we extend the choice of regional distributions for MSNST, and express the statistics for the different definitions of dissimilarity measure. We claim and demonstrate that the choice of regional distributions and dissimilarity measures is a nontrivial task which has a deep impact on the texture segmentation. The influences of them are experimentally compared and analyzed based on the k-means clustering method and Graph Cuts framework. Experiments using a large number of synthesized texture images and real natural scene images demonstrate the superior segmentation performance of Gaussian Mixture Model (GMM) distribution with Riemannian measure.