Unsupervised SAR image segmentation method based on MAP classification criterion and anisotropic diffusion smoothing

摘要

Segmentation is of great importance in the community of synthetic aperture radar (SAR) imaging interpreting and understanding. In this paper we realize an unsupervised SAR image segmentation system based on statistical maximum a posterior (MAP) classification criterion and physical heat diffusion derived anisotropic smoothing process. Generalized mixed Gaussian distribution is applied to model the image gradation with expectation maximize (EM) method implementing the parameter estimation. A novel idea is proposed to linearly combine the Gaussian branch related posterior probabilities to fit the segmentation problem size and this endows cursory initial segmentation robust adaption to a wide range of SAR data variability. Proper use of anisotropic diffusion (AD) on the posterior probability domain can effectively remove the multiplicative speckle noise of raw data and has advantage to smooth the inner area while well preserve region edges, just as optimal ultimate segmentation process requires. A brief introduction of the method is presented along with many application considerations. The correctness and efficiency of the method have been verified by several examples.