An algorithm based on ABNL-MRF model for SAR image segmentation

Abstract

In this letter, a segmentation method is proposed based on an adaptive Bayesian non-local MRF (ABNL-MRF) model. Aimed at the multiplicative speckle existing in SAR images, a patch-similarity measure based on a ratio of probability is introduced firstly. Then non-local similar pixel-blocks are adopted to guide the segmentation of the noisy image. However, non-local method over-smoothes edge regions and makes them inaccurate though it is robust for speckle. A rectification index is designed according to the homogeneity degree of local window. Based on the value of rectification index, a non-local constraint term is adaptively integrated into the potential function. Thus a suitable prior is defined that guarantees both the smoothness of the denoised image and the preservation of its structure. Experimental results demonstrate the very good performance of the proposed method.