SAR Image Target Recognition Based On Robust Bayesian SVM

As a high-resolution imaging radar,Synthetic Aperture Radar(SAR)can detect the targets on the ground without being affected by sunlight and weather precisely and persistently.In addition,SAR can recognize the fake objects and penetrate the occlusions.These advantages make SAR wildly used in the military field.With the acquisitions of large amount of high-resolution SAR images,the SAR image processing such as detection,discrimination and recognition have been developed rapidly.In practice,due to the complexity of background and target configuration as well as the lack of enough training target variants,it is hard to recognize the testing targets.Therefore,how to recognize the target variants becomes an important problem in SAR target recognition.This thesis mainly studies the SAR image preprocessing,support vector machine(SVM),kernel learning method,dropout training method and Bayesian SVM,and proposes a target recognition classifier based on the Bayesian SVM and the multiple kernel learning method.The content of this thesis can be summarized as follows:1.This thesis introduces the SAR image preprocessing method,including segmentation processing,registration processing,cropping processing and filtering processing.2.This thesis simply introduces the dataset(Moving and Stationary Target Acquisition and Recognition,MSTAR)used in the experiments and the experimental scenarios.3.This thesis presents the knowledge of SVM,including the inference and the solution of the maximum margin classifier and the generalized maximum margin classifier,and then Bayesian SVM is presented on the basis of the generalized maximum margin classifier.This thesis presents the detailed inference of Bayesian SVM model,the solution of which is obtained by the expectation maximization(EM)algorithm.In order to verify the performance of the Bayesian SVM,MSTAR dataset is used in the experiment.4.This thesis presents the concept of kernel method and the related content of the single kernel learning method,and then the multiple kernel learning method is studied.Furthermore,the Bayesian SVM based on multiple kernel learning model is proposed.The property of this classifier is verified in the experimental verification stage.In the specific experiment,the radial basis function is employed as the kernel function.The three feature subsets of the MSTAR dataset are the amplitude feature,the frequency domain feature and the sparse coefficient feature based on KSVD and orthogonal matching pursuit(OMP).The experimental results of other methods are also obtained to verify the improvement on Bayesian SVM based on multiple kernel learning.