A Hierarchical Triplet Markov Fields Model For SAR Image Segmentation

As an active detector, Synthetic Aperture Radar(SAR) can provide high-resolution images in all-weather, all day and all night, and has been widely used in military reconnaissance and civil field. The segmentation of SAR image is an important part of target recognition based on SAR image, which can reflect natural attributes and spatial structure of an image, hence more essences of image can be revealed. But because of the backscatters, SAR image are contaminated by speckle noises which will lower image quality and mask image structure. Thus it's an important and difficult problem how to smooth noises and segment targets accurately in the segmentation of SAR images.Markov random field(MRF) algorithm is a popular method for image segmentation since its great ability to describe spatial dependences of pixels based on a neighborhood system of images. However, because it is assumed that each region has homogeneous and uniform properties, the MRF is not suitable to deal with the SAR images containing plenty of texture information. With the introduction of an auxiliary field, triplet Markov random fields(TMF) model makes up for this flaw perfectly to model the non-stationary property of SAR image, which has obtained satisfactory results in the application of SAR image segmentation. However, as a non-causal model, the TMF is supposed to perform better in effective utilization of global information.With the analysis of TMF model, we propose a casual nonlinear diffusion-based hierarchical triplet Markov random fields(ND-HTMF) model for SAR image segmentation. In fact, the proposed method combines superiorities of both TMF and the nonlinear diffusion filter. With the intra- and inter-scale dependences captured by the ND-HTMF model, we are able to extract the global and local characteristics of the image content more effectively. At first, the use of a multi-scale image structure, based on a nonlinear diffusion scheme characterized by noise removal and a good localization of the boundaries of image, overcomes the geometric distortions of most pyramidal approaches. Moreover, the new interpretation and redefinition of the auxiliary field has been implemented. It explicitly distinguishes each pixel as edge or non-edge to offer an additional guidance of boundary localization in segmentation. Lastly, with modeling of the multi-scale likelihood term and the causal energy function through two computation processes separately, the ND-HTMF can acquire a higher accuracy in capturing the global and local image structure information across scales. Experiments with several simulated and real SAR images demonstrate that the proposed method achieves satisfactory improvements in both boundary location and region homogeneity.