Unsupervised Multi-class Segmentation Of SAR Images Using Triplet Markov Fields Models Based On Edge Penalty

Synthetic Aperture Rader(SAR) image segmentation is an important stage for SAR images'recognition and understanding, and the research for the SAR image segmentation algorithm has been a hot spot. However, according to the characteristics of SAR side-view imaging and coherent imaging, the image contains a large number of multiplicative speckle noise, signal to noise ratio is low. These present problems for standard image processing techniques.In this dissertation, we propose a new unsupervised multi-class segmentation of SAR images using the triplet Markov fields(TMF)models based on edge penalty, the new algorithm fuse the traditional energy function of TMF model with the principle of edge penalty, which could prevent segment from smoothing across boundaries while modeling the non-stationary SAR image better. Then we optimize the objective function that stems from the new energy function to obtain an iterative multi-region combined Bayesian maximum posteriori model (MPM) segmentation equation for the new segmentation algorithm. Simulated data and real SAR images are applied to evaluate the performance of the proposed algorithm in segmentation. Experimental results and analysis indicate that compared with the classical Markov random field (MRF) and the recent TMF segmentation algorithm, the proposed algorithm effectively improves the segmentation accuracy of the SAR image while reducing the influence of multiplicative speckle noise, with the weak edge location being more accurate especially.