Priors and constraints in Bayesian image segmentation based on finite mixtures

The use of prior densities for image segmentation within the framework of finite mixture models is investigated. Segmentation is posed as a pixel labeling problem and a generalized expectation maximization algorithm for the Bayesian estimation of the pixel labels is employed. This algorithm is based on a unique spatially-variant mixture model that has the flexibility of incorporating any useful prior information on the potential label configurations. Two different priors are proposed for pixel labeling and their effectiveness is assessed quantitatively on simulated images at various noise levels. A qualitative evaluation is also presented using clinical magnetic resonance images of the human brain.