Supervised SAR Image MAP Segmentation Based on Region-based Hierarchical Model

A supervised multi-resolution segmentation method for synthetic aperture radar (SAR) images is proposed. In the method, a new region-based half tree hierarchical MRF model is proposed for multi-resolution segmentation. The watershed algorithm is applied in every resolution level to obtain our region-based model except for the coarsest level. The nodes of the region-based model are defined as regions instead of pixels and region probability is propagated between layers. We redefine Bottom-up pass and Top-down pass in the region-based model and deduce the region-based upward and downward MAP estimation. The segmentation at the coarsest level is obtained with the noncausal MRF model. At the finer levels, segmentation is determined by the probabilities of regions which are produced by watershed algorithms. The experiments demonstrate that the proposed method performs better than the pixel-based hierarchical model as well as the Gibbs sampler with the single resolution model.