Unsupervised SAR Image Segmentation Based On Triplet Markov Fields

Synthetic aperture radar(SAR) systems find increasingly wide applications in both civil and military fields since they can operate days and nights and under any weather conditions. SAR image segmentation is a key and fundamental part of target recognition and interpretation of SAR images. In recent years, this area has gradually become a research hotspot. Nevertheless, due to coherent illumination, SAR images suffer from strong speckle noise, which makes SAR image segmentation a challenging task. Therefore,the research on SAR image segmentation is of great significance for the development of SAR system.Triplet Markov fields(TMF) model has the ability to consider the non-stationary property of images, and can adopt various statistical models to accurately model data. Consequently, it is suitable to deal with non-stationary SAR image segmentation. In order to realize real-time, robust, and effective segmentation of SAR images, based on satellite-borne and airborne SAR images, this dissertation studies unsupervised SAR image segmentation methods based on TMF model, combing with Bayes theory, graph theory, scale space, graph cut and belief propagation theory. The main contents of this dissertation are summarized as follows:(1)An unsupervised SAR image segmentation algorithm using TMF model with edge location is proposed. First for the statistic characteristics of multiplicative speckle noise in SAR image, an edge strength based on ratio of exponentially weighted averages(ROEWA) operator is introduced into the Turbopixels algorithm to obtain a superpixel graph with accurate edge location in SAR image. To enhance the computational efficiency and suppress the speckle, the TMF model on pixel is generalized to that on superpixel graph. Then the new corresponding potential energy function and maximization of posterior marginal(MPM) segmentation formula are derived. The experimental results on synthetic and real SAR images show that the proposed algorithm can obtain accurate edge location in multi-class segmentation of SAR images, as well as enhance the computational efficiency.(2)An unsupervised SAR image segmentation algorithm using TMF model defined on multiscale region adjacency tree is proposed. The noncausal nature of the traditional TMF model leads to iterative inference algorithms that are computationally demanding. Thus, we propose a causal triplet Markov fields model defined on the hierarchy of a multiscale region adjacency tree. The multiscale representation of SAR image is generated by speckle reducing anisotropic diffusion(SRAD), which leads to speckle reduction and good localization of boundaries. The gradient watershed algorithm is used to detect regions and build parent–child linking of regions between scales. The image segmentation is then treated as a hierarchical TMF labeling problem using a noniterative estimation of the modes of posterior marginal(MPM). Experiments on simulated and real SAR images demonstrate the effectiveness of the proposed model compared to the traditional noncausal TMF model.(3)An unsupervised SAR image segmentation algorithm based on TMF model with belief propagation is proposed. Traditional optimization of TMF model often uses complex sampling technology, which is slow to converge, and not guaranteed to converge to the true values. Focusing on efficient statistical inference of TMF model, for the two label fields in TMF, belief propagation algorithm is generalized to the bivariate case to estimate the joint posterior marginal probability of the two label fields through message passing. The two label fields can be simultaneously estimated according to MPM criterion. Experiments on both simulated and real SAR images demonstrate that the proposed algorithm can efficiently suppress the influence of speckle, and obtain accurate segmentation results with reasonable computational cost.(4)An unsupervised SAR image segmentation algorithm based on TMF model with graph cuts is proposed. Considering the existence of two label fields in TMF model, the proposed algorithm iteratively estimates one label field with the other fixed until converges. Graph cuts algorithm is used to find the optimum estimation of each label field under the criterion of maximum a posteriori(MAP). Experiments on simulated and real SAR images demonstrate that the proposed algorithm combines the ability of TMF to accurately model non-stationary SAR images and the efficiency of graph cuts in inference, and can obtain much smoother and more accurate segmentation results with smaller computational cost.