Sar Automatic Target Recognition Method Based On Manifold Learning

Synthetic Aperture Radar (SAR) has the advantages of all-weather, wide, strong penetrable ability and could supply detailed ground mapping material and images in atrocious weather with high resolution, which provides us with a more reliable means of the target recognition. SAR automatic target recognition (ATR) plays a more and more important role in the battlefield awareness, the research and development of SAR automatic target recognition is a hot spot competitively accordingly. Manifold learning is a new type of machine learning theory developed recently, which aims to find the internal distribution of the high-dimensional dataset. It helps to identify the intrinsic characteristics of SAR image dataset, and improve the recognition performance of the SAR ATR system.In this dissertation, SAR image preprocessing, feature extraction and recognition are studied, and the main contents are as follows:1. For the requirements of SAR Image preprocessing algorithm to preserve detailed information and improve the discriminative power of the preprocessed images, the complex preprocessing algorithm of combining the image segmentation algorithm of two-parameter CFAR image segmentation based on weibull distribution - geometric cluster - binary mask and the image post-processing algorithm is investigated, which could maintain the edge, texture and other details of the target effectively, and improve the discriminative power.2. For the traditional feature extraction algorithms based on global linear structure can not effectively overcome the nonlinear effect of the high-dimensional SAR dataset, Generalized Neighbor Discriminant Embedding (GNDE) is proposed for feature extraction, including GNDE on Vector (GNDEV) and GNDE on Tensor (GNDET), which can solve the nonlinear problem of SAR dataset effectively, and improve the discriminative power of the extracted feature and the robustness to the datasets. So GNDE can improve the recognition rate significantly.3. For a single feature extraction method always can not meet the recognition rate requirements, 2DPCA-based GNDET is explored for SAR image feature fusion, which can both compresse the dimensions of the feature matrix and enhance the discriminative power of the extracted feature.4. For the traditional feature extracton methods of manifold learning always result in large computation load, Scaling Discriminant Embedding is proposed for feature extraction, including SDE on Vector (SDEV) and SDE on Tensor (SDET), which could enhance the discriminative power of the extracted feature at low computation load, so it can improve the recognition rate effectively.