Synthetic Aperture Radar(SAR) has the ability of all weather, day/night imaging, and is an important method for earth observation. SAR target recognition which uses SAR image information to recognize the target class and type. There is a specific application requirements for SAR target recognition in battlefield reconnaissance, precision attack and other military areas. SAR target recognition can enhance the SAR sensor ability on information perception, and is one of the key technologies to implement the application of SAR technology.In recent years, substantial progress has been obtianed in the related theory and methods research on SAR target recognition technology research. However, in the fields of the target database construction, target feature database establishment, rapid detection and location, classification and decision, there are still some problems, like the theoretical research is not deep enough, the image information can’t be used rapidly, the recognition method is lack of practicability, and the intelligent level of the recognition system is low, which limits the further development of the SAR technology.The work in this thesis focuses on target feature database establishment, classification and decision, which are the key problems in SAR target recognition. The main work includes sparse feature exaction, multi-feature multi-classifiers fusion, and intelligent learning.The major innovations in this thesis are summarized as following:1. A novel variant of Non-negative Matrix Factorization(NMF), called L1/2-NMF is proposed, and is carried out a thorough study by applying it in SAR target recognition. The L1/2 norm constraint is introduced to NMF, which can increase the sparseness of the base and feature matrixes, reduce the feature information redundancy. L1/2-NMF can improve the performance of NMF on target feature description;2. A hierarchical propelled fusiuon strategy for classification decision, which has the advantage of parallel and cascade fusion structure, is proposed in this thesis. The multi-feature can be fused effectively, and the features conflict can be reduced by the way of hierarchical decision and hierarchical fusion. This method can improve the efficiency of multi classifier fusion, and use the multiple features effectively;3. To solve the small sample data set recognition problems, a recognition method based on hierarchical representation of the target feature is proposed. The hierarchical target features representation can be obtained by muilti-layer network model of Deep Learning alogithm. The best hierarchical feature is choosen to feed to the pattern classifiers, which can improve the target recognition accuracy rate on the small sample data set;4. A SAR target recognition method based on the Constrained Restricted Boltzmann Machine(CRBM) is proposed. According to the sparse representation problem of the features obtained by deep structure, the generalized sparse constrain is introduced to RBM, which can enhance the sparseness of base matrix and hideen units, to improve the effectiveness of target feature representation by RBM. Meanwhile, the CRBM is introduced to the hierarchical feature recognition method, which can significantly improve the performance of SAR target recognition..All the above works have been verified by MSTAR public database, that they can improve the effectiveness and accuracy of SAR target recognition. |