The Research On Parameters Estimation Of The Generalized Gamma Mixture Model And Application On Sar Image

SAR (Synthetic Aperture Radar) is widely used in field of military and civil, the accurate statistical modeling of SAR images is a crucial problem in the context of effective SAR image processing and applications.(Finite Mixture Model) has been widely used in statistics field for its flexible and strong description capability. As a powerful probabilistic modeling tool for univariate and multivariate data, its applying even extended to biology psychology and gene field. Because of simple form, convenient calculation, GMM (Gaussian Mixture Model) has become widely used in the Mixture Model. However, with the development of computer technology, problems we faced during our research becomes more and more complicated, GMM can not fully describe and characterize these complex data considering that actual data has nonlinear, non-Gaussian characteristics. In this paper, we will focus on the study of GGMM (Generalized Gamma Mixture Model) which has stronger ability of description, Discussing about how to use this model fitting data and get the estimate parameters.The most important issues in mixture modeling include the selection of the number of components and parameter estimation. The usual choice for obtaining estimates of the mixture parameters is the EM (Expectation Maximum) algorithm, but EM is highly dependent on initialization and can hardly converge to the global optimal if it get a bad start point. This paper will hybrid the PSO (Particle Swarm Optimization) algorithm with EM algorithm by using PSO algorithm to find the general location of global optimal and use it on the start point of EM algorithm. EM algorithm can accelerate convergence to global optimal because of this right initial value. After the formulas of EM algorithm on estimation of GGMM are deduced, we find out the parameter settings which get the best performance of PSO algorithm applying for GGMM estimation through large number of simulation. Simulation compared with the random initiation EM algorithm proved the feasibility of PSO combining with EM algorithm.MML (Minimum message length) criterion is chosen to solve the problem of selecting the number of components. It will choose the number smaller than the truth while other criterion will choose the bigger number when get the wrong result. Through further study of this criterion, the MML expressions of selecting the number of GGMM carried out under the hypothesis of prior of parameters. MML Combined with the hybrid of PSO and EM algorithm, applying to the model selection of GGMM as the complete algorithm. Simulation with various data verified the good performance of MML criterion on selecting the number of components, and the whole algorithm can choose a highly fitting model. Selecting a correct model for SAR image certificate it provided the theoretical foundation for practical application.