Synthetic aperture radar image segmentation based on improved fuzzy Markov random field model

Fuzzy Markov random field (FMRF) is a novel model for data clustering or image segmentation. This paper presents an improved FMRF segmentation method for synthetic aperture radar (SAR) images. The originality of this algorithm based on the fact that there are many interlaced edges or mixed areas among different types of regions in SAR images, which are often caused by the limited resolutions of the acquisition systems or speckle noise of sensor. Fuzzy method is just a shortcut for this kind of uncertain classification, which could give more flexible and reasonable segmentation than the `hard´ method. Although many scholars had done much work in the two-level FMRF segmentation, we improved the algorithm from two-level to multi-level and gave more clear and reasonable threshold definitions for hard and fuzzy MRF. The first part of our work involves definitions of the multi-level FMRF and standards for hard and fuzzy MRF. Then we apply the simulated annealing (SA) and expectation-maximization (EM) algorithms to search global optimal resolutions and estimate unknown parameters. Finally the segmentation experiments of two SAR images demonstrate that the proposed algorithm is efficient to distinguish interlaced edges or mixed areas and successful to restrain noise