Research On Polarization SAR Target Detection And Tracking Algorithm Based On Fully Convolutional Network And Transfer

SAR radar has developed considerably after years of research,and the interpretation of SAR images has also made great progress.However,with time going by,the traditional singlepolarization SAR images can no longer meet the current increasing needs,such as SAR images cannot provide more physical information of the target area,traditional airborne SAR and satellite-based SAR cannot provide continuous and long time observation of specific target areas.Therefore,two modes of polarized SAR and video SAR have been developed on the basis of SAR imaging.The research of this thesis is mainly about image interpretation of the two newly emerged polarized SAR and video SAR modes,which can be roughly divided into three parts: basics of polarization and convolutional neural network,ship target detection method for polarized SAR images,and moving target tracking method for video SAR.the major work of this thesis is as follows.1.The theoretical basis of polarization SAR and convolutional neural network is introduced.In the polarization SAR part,a simple theoretical derivation is made in terms of the causes of the polarization mode of electromagnetic wave generation,the characterization mode of polarization features and several traditional polarization decomposition methods.In the convolutional neural network section,the foundation of convolutional networks is introduced,then several common neural network structures are introduced,and finally two important convolutional neural networks are introduced,besides the application of migration learning in small sample conditions is presented.2.An end-to-end polarized SAR ship target detection network is proposed.For there is no open source polarized SAR target detection dataset,a small polarized SAR ship target detection dataset is constructed in this thesis.For the current research methods of polarized SAR target detection,which usually use the traditional polarization decomposition algorithm,this thesis designs a fully convolutional neural network structure to replace the traditional polarization decomposition algorithm and connects with the Faster R-CNN network to form an end-to-end polarized SAR ship target detection network.In this thesis,several experimental control groups are set up,and the polarized SAR ship target detection dataset constructed in this thesis is used to train and test the proposed network and the control group network,and finally the effectiveness of the proposed network is verified by COCO evaluation index.3.A video SAR moving target tracking algorithm based on target detection and Kalman filtering is proposed.At present,video SAR generally uses a higher frequency carrier frequency,which makes the moving target imaging in video SAR will have a very large offset and scatter,and thus will leave a shadow at the real location where the target is located.In the absence of sufficient video SAR data,it is difficult to use migration learning to apply existing target tracking networks to video SAR,for which a target tracking method based on the Kalman filter algorithm for Faster R-CNN shadow target detection is proposed in this thesis.Then,the algorithm in this thesis is verified on a video SAR published by Sandia Lab.