SAR Automatic Target Recognition Research Based On Neighborhood Discriminating Embedding

As an important aspect of information synthesis processing in battlefield, SyntheticAperture Radar Automatic Target Recognition (SAR ATR) is provided with variouskinds of academic value and wide application prospects in military, which has becomean important research topic currently.Based on the process of SAR ATR system, this dissertation covers two researchaspects which includes SAR image preprocessing and feature extraction of targets. Themain contents of this thesis are as follows.(1) Due to the problems of the low proportion of targets in the original SARimages and the SAR images with high dimension, two-parameter CFAR based onWeibull distribution and geometric clustering method are utilized to extract the Regionof Interest (ROI) and reduce the dimension of the original SAR images. Besides, wetake advantage of the image post-processing techniques, including centering imageregistration, energy normalization and gray enhancement based on power function, toenhance the recognition information from the extracted targets.(2) For the issue of traditional algorithms based on global linear structureextracting feature from high-dimensional images inefficiently, manifold learning isintroduced in SAR image feature extraction. We validate the manifold distribution ofhigh-dimensional SAR images with both theoretical analysis and experimentalverification, which lays the foundation for the manifold learning applied in SAR ATR.Meanwhile, the concept of SAR ATR based on neighborhood discriminating embeddingis proposed and illustrated in our dissertation.(3) Because the traditional manifold learning algorithms can not explore theclassification information from the samples’ neighborhoods, Neighborhood VirtualPoints Discriminant Embedding (NVPDE) is proposed based on the concept ofneighborhood discriminating embedding. By introducing the neighborhood virtualpoints in each sample’ neighborhood, the spatial relationships between theneighborhoods are established, which is able to improve the recognition rate of theextracted feature. The best recognition rate of NVPDE method is97.88%. (4) In order to solve the problem that the ambiguous clustering direction of thesamples in the feature extraction based on traditional manifold learning algorithms,Neighborhood Geometric Center Scaling Embedding is proposed based onneighborhood discriminating embedding, including One-dimensional NeighborhoodGeometric Center Scaling Embedding (NGCSE), Two-dimensional NeighborhoodGeometric Center Scaling Embedding (2DNGCSE) and Two-dimensional PrincipalComponent Analysis based Two-dimensional Neighborhood Geometric Center ScalingEmbedding (2DPCA-based2DNGCSE). In this algorithm, neighborhood geometriccenter scaling for each sample is established, which is able to endow each sample withclear clustering direction in the feature extraction. The extracted feature will beclassified effectively in the low-dimensional space, and the recognition rate and stabilityof the algorithm are improved. The best recognition rates of those three methods are97.88%,97.95%and98.24%respectively.The experiments based on MSTAR database have been conducted and verified theeffectiveness of the SAR image preprocessing and feature extraction methods in ourdissertation.