Hidden multiresolution random fields and their application to image segmentation

In this paper a new class of random field, defined on a multiresolution array structure, is described. Some of the fundamental statistical properties of the model are established. Estimation from noisy data is then considered and a new procedure, multiresolution maximum a posteriori estimation, is defined. These ideas are then applied to the problem of segmenting images containing a number of regions. Implementation of the Bayesian approach is based on a multiresolution form of Gibbs sampling. It is shown that the model forms an excellent basis for the segmentation of such images, which works with no a priori information on the number or sizes of the regions