Study On SAR Images Target Recognition Based On Feature Sparse Representation

In recent years, lots of attention is paid to sparse representation theory, which has beensuccessfully used in image compression and denoising. Target recognition in SAR images is todetermine targets’ property, class or type. Target recognition technology based on SAR images playsan important role in military, civil and other areas. Supported by aviation fund project, this paperrealizes SAR image target recognition using sparse representation by tracking domestic andinternational research achievement. The completed major tasks are as follows:(1) This paper studies one of the sparse representation theory contents: the dictionary. First westudy the development of a fixed dictionary and methods of dictionary learning. Then we analyze theapplication of dictionary in signal processing, such as images and etc, and sum up three principlesfor over-complete dictionary design in signal sparse representation.(2) This paper studies the other content of sparse representation theory: sparse algorithm. Weanalyze the characteristics of matching pursuit algorithms,l1norm regularization based algorithmsand iterative shrinkage algorithms, and propose an improved orthogonal matching pursuit algorithm.The new algorithm determines candidate atoms adaptively via nonlinearly decreasing threshold,excludes the atoms with smaller energy through regulazation, then reconstructs signal according toleast squares method. Simulation results show that the proposed algorithm has higher signalreconstruction performance, which is faster than BP algorithm and more accurate than matchingpursuit algorithms.(3) This paper studies application of sparse representation theory in SAR images targetrecognition. There is a high dimensional problem here based on sparse representation in pixeldomain. To overcome this problem, we extract the low-dimension and high-precision G2DPCAfeature to establish an over-complete dictionary, and learn the dictionary according to lineardiscriminant principle, which cuts down complexity in sparse solving greatly. Then sparserepresentation coefficient of test sample is computed based on the studied dictionary. Classificationand recognition is realized according to SCI value of coefficient. Simulation results show that, thismethod achieves both high accuracy and high speed in recognition with only simple preprocessing.