Multiscale MRF-based texture segmentation of SAR image

We propose a multiscale Bayesian segmentation algorithm for SAR image in this paper. A hierarchical two-level Markov random field (MRF) is applied to represent both texture and region label over the wavelet lattice. The high level uses an isotropic Multi-level logistic (MLL) random field to characterize the blob-like region formation process at each scale and the interscale dependencies over the corresponding multiresolution region. At lower level a novel Causal Gaussian autoregressive (CGAR) process is proposed to describe the fill-in of multiresolution region. Once the multiscale double MRFs model is established, in term of Sequential maximum a posteriori (SMAP), model parameter estimate and region segmentation are performed alternately from coarse to fine scale. Our segmentation method is tested on both synthetic and ERS-1 SAR images.