Study On SAR Target Recognition Based On Sparse Representation

Synthetic aperture radar(SAR) is an active, all-weather remote sensor that can run during the day and night. It can also penetrate clouds and construct high resolution images of the groud from airbone platform which plays a very important role in modern battlefield.Automatic target recognition(ATR) based on SAR images is of great importance in the military field and has been a hot research topic.In recent years, the proposed of compressed sensing provides a soil of engineering applications for the development of the sparse representation which has been applied to many fields such as image compression, denoising and recognition. Experiments on face database show that sparse representation method can obtain better results than traditional methods. On this basis, sparse representation is applied to SAR target recognition. In this paper, the basic theory of sparse representation is introduced, including the construction of sparse dictionary and sparse algorithm. Based on this, SAR target recognition combined Kernel Principal Component Analysis(KPCA) with sparse representation is proposed. First, KPCA feature extraction is used to get the feature of the samples. Then a sparse representation model is built in the feature space. The sparse coefficient is obtained by GPSR(Gradient Projection for Sparse Reconstruction). Finally, the recognition is achieved by computing the energy of the sparse coefficient. Experimental results with MSTAR(Moving and Stationary Target Acquisition and Recognition) SAR data sets show that the proposed method can improve the target recognition result without knowing the target azimuth. And the proposed method is an effective method for SAR target recognition.In addition, this paper studies a new feature extraction method namely sparsity preserving projections(SPP), which introduces the sparse coefficient to the feature extraction and gets the feature vectors through the relationship of sparse reconstruction. On the basis of this method, an improved SPP feature extraction is proposed. This method can maintain the relationship of sparse reconstruction between samples and draw on the idea of locality preserving projection(LPP) feature extraction, which makes the extracted features not only keep the sparse reconstruction character, but also make the distance between the same samples smaller. The improved method is applied to SAR target recognition with the MSTARdata. The experimental results demonstrate the effectiveness of this method.