Low-level bayesian segmentation of piecewise-homogeneous noisy and textured images

We present a novel approach to image segmentation, differing from the known “simulated annealing” method in the following ways: the compound Bayesian decision rule and consequent maximal marginal a posteriori probability (MMAP) estimates of desired region labels in pixels; the two- or three-level piecewise-homogeneous Gibbs random field with constant control parameters as the probabilistic model of the images and region maps (in the general case such a model integrates the submodels of the region map, of the ideal intensities within each region, and of the noise distorting the ideal intensities); the stochastic relaxation with the constant control parameters of the Gibbs probability distribution only as a tool to obtain the samples of this field and estimate the unknown marginal a posteriori probabilities of the region labels by collecting in each pixel the histogram of labels for these samples; the like stochastic relaxation with directed variation of the control parameters of the Gibbs probability distribution as a tool to find maximal likelihood estimates of the unknown these parameters. Some experimental results are presented.