Multiscale random field models for Bayesian image segmentation using a sequential maximum a posteriori (SMAP) estimator was also considered by our group, but was not implemented. This SMAP approach uses a multiscale label field model in which regions of varying size are assigned labels corresponding to the region type. The progression from coarse to fine scale of these regions or fields is assumed to form a Markov chain. This implies then that the distribution of the labels at any scale given all coarser scales depends only upon the scale that is one step coarser. The SMAP criteria uses a cost function that assigns progressively greater cost to segmentations with larger regions of missclassified pixels. This tends to yield better results than a traditional MAP criteria which minimizes the probability that any single pixel is missclassified. The SMAP method minimizes the spatial size of errors, tending to yield subjectively better segmentations.
The actual algorithm would require using a set of training data to determine parameters for a Gaussian mixture distribution. These parameters can then be used to segment the image.
This SMAP approach required a priori information about the regions that are intended to be segmented. Since we were more interested in developing an algorithm that segmented a general image without any notion of what the regions might be, this technique was not very relevant.