Evolutionary Gibbs sampler for image segmentation

We propose a novel evolutionary algorithm for the function optimization problem in Bayesian image segmentation with Markov random field prior. Function variables are partitioned into several codings. A pivot coding is selected and variables in it are evolved respectively according to their probability distributions which encode both the evolutionary pressure and contextual constraints from neighboring pixels. Variables in other codings are evolved according to their conditional probabilities. In summary, the algorithm is about building probabilistic models to guide search. It achieves the efficiency and flexibility by incorporating Gibbs sampler in an evolutionary approach. Remarkable performance is observed in some experiments.