Unsupervised Mumford-Shah energy based hybrid of texture and nontexture image segmentation

In this paper, we extend Mumford-Shah energy, the most general model for image segmentation on the assumption that all visible surface patches have intensity which is slowly varying plus noise, to unsupervised texture segmentation by considering multiple channels including both colors and Gabor wavelets representations. The Mumford-Shah energy in the paper thus formalizes a tradeoff between region homogeneity in color and texture field and edge compactness. Numerical results are presented for synthesized and real images to reveal the validity of the proposed algorithm.