Feature Extraction Technique In SAR Automatic Target Recognition

Synthetic Aperture Radar(SAR) Automatic Target Recognition(ATR) technology has very important applications in military and civilian fields, becoming a hot research topic in the world. Feature extraction, as a key issue in SAR ATR, mainly focuses on how to transform the samples from high dimensional space to low one and extract the most effective characteristic features.SAR ATR feature extraction based on projection space is deeply studied in this thesis. A linear optimal projection space method is proposed, to construct a projection space having the most similarity with own value space and absolute orthogonality with other ones. Thus, the similarity between different target spaces is greatly eliminated and the separability between different targets is widely enlarged. With the analyzing on the relationship between azimuth and SAR target distribution, another feature extraction method based on hyper-sphere projection space is proposed. By this, SAR samples'distribution characteristics of all-round are obtained and the affect of azimuth on SAR target recognition is significantly reduced.Moving and Stationary Target Acquisition and Recognition(MSTAR) dataset from American Defense Advanced Research Projects Agency(DARPA) and Air Force Research Laboratory(AFRL) is used for experiments and verification, from which the efficiency of both methods has been conformed.