Research On SAR Target Recognition Algorithms Based On Sparse Representation

Synthetic Aperture Radar(SAR) which is one kind of high resolution imaging radar is widely applied in civilian and military fields. Since SAR target recognition which is one of the applications is important for defense warning, it has become one hot research topic. Sparse representation selects a few atoms from the over-complete dictionary to linearly reconstruct signals. When sparse representation tackles recognition problem, two benefits are shown. One is that natural discriminative information will be contained in the sparse representation coefficients, and the other one is that the recognition results show good immunity to noise which is a challenge in SAR target recognition. Hence, it is promising to study SAR target recognition based on sparse representation theory.In this context, the researches on SAR image preprocessing, feature extraction and target recognition based on sparse representation are carried in this paper. Here are the main contents:1. The SAR image preprocessing based on SVM(Support Vector Machine) is introduced with consideration of the characteristics of the SAR images in MSTAR database. The details of the SAR target are kept through logarithmic transformation, stable segmentation based on SVM and post-processing to SAR images. Meanwhile, the speckle noises in SAR image are reduced in the preprocessing procedures. A sub-image with a clear indication of the SAR target is then generated after the above procedures for a better recognition.2. Maximize Sparse Reconstruction Margin Projections(MSRMP) whose sparse representation model better reflects the discrimination is proposed to improve the Sparse Neighborhood Preserving Embedding(SNPE). MSRMP enhances the recognition performance and shows robustness to classification methods. When the number of dimensionality is enough, even if SAR image has dense noises, the recognition rate will be stable whatever the classification strategy is.3. Since the SAR images under multiple views of one SAR target can be obtained in real scenario, a discussion about the Joint Sparse Representation(JSR) model is conducted to combine the SAR target information from different views. An Improved Joint Sparse Representation(IJSR) model is proposed on the foundation of the JSR model. A common pattern is sought through the L1-norm minimization model and the low-rank matrix recovery method in the IJSR model firstly, and then the label of the inputting SAR target is inferred according to the information extracted from the common pattern. With the sparse representation classification strategy, a better recognition performance of the IJSR model on the MSTAR database is shown compared with the recognition performance of the JSR model.