Target Recognition Of SAR Images Based On Sparse Representation

The target recognition of SAR image plays a key role of information acquisition and has important application value, which has been a hot research topic in the field of target recognition at home and abroad. In recent years, the sparse representation theory is applied to all kinds of image processing field, which obtains good recognition performance in the field of face recognition. Based on the basic theory of sparse representation, this thesis concentrates on the two key steps of SAR image target recognition: the structure of the redundant dictionary and the solution of the sparse coefficient. The main contents are as follows:(1) To overcome the weakness of low class difference and large scale in the original redundant dictionary, the novel method of bidirectional dictionary compression is proposed according to the two-dimensional redundant dictionary. Longitudinal improvement of dictionary: the certainty feature of low-frequency image in wavelet domain is extracted by 2DPCA, since there is a feature in SAR image that is mainly composed of certainty information and uncertainty information. Transverse improvement of dictionary: based on K-NN algorithm, the dynamic dictionary of neighbor subspace is obtained by filtering the dictionary atoms dynamically.(2) The sparse decomposition algorithm based on redundant dictionary is studied. By the contrast of L1 norm minimum convex optimization algorithm and OMP algorithm, a conclusion is drawn that later algorithm in the difference of the coefficient histogram distribution and decomposition efficiency is both better than the previous algorithm. To solve the problem of the sparse degree K unknown, we replace K with the category statistics C as the iterative termination condition, and the simulation result shows that the modified OMP algorithm can achieve better recognition rate.(3) According to the distribution characteristics of sparse decomposition coefficient, the maximum coefficient criteria and the classification coefficient maximum criteria are concluded. By the contrast of two the recognition rate under the two classification criteria, the simulation result shows that the classification coefficient maximum criterion can achieve better recognition rate. So, we designed the classifier according to the criteria.(4) Based on MSTAR database, the statistic recognition result shows the algorithm robustness in case of some images contained noise, shade and resolution reduction.