Image segmentation with a class-adaptive spatially constrained mixture model

We propose a hierarchical and spatially variant mixture model for image segmentation where the pixel labels are random variables. Distinct smoothness priors are imposed on the label probabilities and the model parameters are computed in closed form through maximum a posteriori (MAP) estimation. More specifically, we propose a new prior for the label probabilities that enforces spatial smoothness of different degree for each cluster. By taking into account spatial information, adjacent pixels are more probable to belong to the same cluster (which is intuitively desirable). Also, all of the model parameters are estimated in closed form from the data. The proposed conducted experiments indicate that our approach compares favorably to both standard and previous spatially constrained mixture model-based segmentation techniques.