Bayesian learning, global competition, and unsupervised image segmentation

A novel approach to unsupervised stochastic model-based image segmentation is presented, and the problems of parameter estimation and image segmentation are formulated as Bayesian learning. In order to draw samples corresponding to different classes, a global competition strategy is adopted for label commitment based on the "power-value" associated with each sample (or site). The smaller the value, the more powerful the sample to compete. Parameter estimation and image segmentation are executed in the same process. Bayesian modeling of images by Markov random fields (MRF) makes it easy to represent the power of each site for competition. The new procedure for unsupervised image segmentation is performed on synthetic and real images to show its success.