Research On Variational Problems In SAR Image Segmentation

The segmentation of Synthetic Aperture Radar(SAR) images is the fundamental and key step in the SAR image processing. Due to the impact of speckle noise, it is necessary to segment the SAR images in the light of the inherent characteristics of SAR images. The method of SAR image segmentation, which is based on the variation theory, can build different energy functionals according to different characteristics of SAR images. The segmentation results can be obtained by minimizing the energy functional through the variational method.This thesis studies the technology of variational segmentation of SAR images. The main contents of this thesis are as follows:(1) The GAC model and the CV model, which are based on the boundary information and region information respectively, have been investigated. And the mixture model has been expanded to the multi-region condition. Because of the time-consuming of the level set method, this thesis proposes a fast dual algorithm which is based on the smooth dual model derived from the Potts model. Comparing the fast dual algorithm and level set method, the experiments show that the fast dual algorithm can get the segmentation results more effectively and quickly.(2) The thesis analyses the Gamma distribution of the uniform SAR images and the texture characteristics of the non-uniform SAR images. The texture characteristics, which can make up a characteristic vector, could be extracted by using the Grayscale Co-occurrence matrix(GLCM). According to the Potts model, the energy functional can be set up by combining the vector and the Gamma distribution. The segmentation results, which are obtained by using the Gamma distribution or the texture characteristics individually, may be less accurately than the segmentation results when the integrated method is used.(3) The thesis utilizes the Potts model and the complex Wishart distribution of the polarimetric coherent matrix to establish the energy functional. This thesis uses the fast dual algorithm to minimize the energy functional. During the minimizing, the H?? classification method and the complex Wishart distribution are used to initializing. This method can get the initialization curve and cluster number automatically. Further, the automatic initialization takes the statistical characteristics and scattering characteristics into consideration simultaneously. The results are much more in line with the real terrain information when comparing with the random initialization.