Research On SAR Images Segmentation Based On Generalized Gamma Mixture Model

SAR image segmentation is to separate the targets of interest in the SAR images for the purpose of precisely analyzing these targets, which has been widely used in many fields such as military reconnaissance, natural resource exploration, maritime surveillance, agro-forestry monitoring, and so on.So far, a large number of methods for image segmentation have been proposed, in which the commonly-used one is the Gaussian-mixture-mode-based technique. However, due to the complexity, non-linear and non-Gaussian of the SAR image data, the traditional Gaussian mixture model (GMM) can not accurately describe the SAR image data. To this end, we will make use of flexible and powerful generalized Gamma mixture model to model the SAR images, and then establish the basis for accurate image segmentation.In this thesis, SAR image segmentation based on generalized Gamma mixture model is researched, whose main content will be stated as follows:(1) Firstly, the minimum-error segmentation method is derived from two-component generalized Gamma mixture model and minimum error criterion. Then, the differential method of histogram is applied to determine the location of valley points and their number. Accordingly, the group number of histogram is deduced from the number of valley points. For each group, the threshold is adaptively obtained by using the proposed minimum-error method based on two-component generalized Gamma mixture model. Finally, we achieve the multi-threshold segmentation of SAR images. Compared with several existing segmentation algorithms, the proposed algorithm is feasible and effective, and can obtain the better segmentation results.(2) Aiming at the limitations of global threshold method and truncation effect caused by the histogram grouping, as well as the complexity of SAR image data, the generalized Gamma mixture model with adaptive number of components is employed to model the SAR images, whose parameters are estimated by CQPSOEM algorithm, and the number of components are determined by using the MDL algorithm. Specifically, the CQPSOEM method can effectively improve the accuracy of parameter estimation along with low computational complexity, by integrating the simplicity of QPSO and the ergodicity of chaotic mechanism. Further combining the Bayesian minimum error rate criteria, we achieve the segmentation of SAR images. By comparison with the previous algorithm and several existing ones, the proposed algorithm is more robust and accurate.