Unsupervised multiresolution image segmentation based on color moments

This paper describes a novel multiresolution image segmentation algorithm which is designed to separate a sharply focused object of interest from background. According to the principle of human vision system, the algorithm first searches salient blocks in global image domain. Then a multiscale approach based on color moments is used to perform context-dependent classification of these blocks. The algorithm is fully automatic in that all parameters are image independent. Unlike other color-based approaches which use global optimization methods, our algorithm does perform a multiresolution process and achieves better segmentation results at higher speed.