SAR Automatic Target Recognition And Related Techniques

SAR (Synthetic Aperture Radar) ATR (Automatic Target Recognition) is crucial to the success of battlefield awareness and has become a very hot research topic. In recent years, radar target recognition has made steady progress in many fields, including feature extraction, target classification and recognition. Some ATR systems have been built and have been successfully used in the areas like ground detecting and precision guidance in spaceborne/airborne SAR.This thesis first reviews the ATR fundamentals and the state-of-the-art development of SAR ATR techniques. Main contributions include:First, a novel image alignment approach is proposed, which is an efficient matching method based on non-linear least square (NLS) fitting. The alignment, amplitude as well as average bias compensation can be done simultaneously in the frequency domain. The new preprocessing method exhibits better performance in matching accuracy and matching speed. Secondly, an optimal linear transform based SAR ATR approach is proposed, which takes optimal linear transform over the magnitude frequency response of target samples and uses the SVM (Support Vector Machine) as the classifier. This method exhibits shift-invariance property and can effectively improve Pcc (Probability of Correct Classification) and training as well as testing speed. Moreover, it can lower down the requirement of bearing estimation accuracy of targets. Thirdly, kernel feature extraction is first applied to SAR ATR. The target feature is extracted by using KPCA (Kernel Principal Component Analysis) or KFD (Kernel Fisher Discriminant) and then classified by SVM. Experimental results demonstrate that the proposed methods can achieve higher Pcc, faster training and testing speed, and better generalization ability. In addition, they are not sensitive to the uncertainty of the target aspect.Finally, two approaches are proposed for SAR target and shadow segmentation. One is a simple segmented method based on Weibull distribution, which shows better performance than Gaussian distribution based methods. The other is an improved method based on MRF (Markov Random Field), which can improve the processing speed while maintaining image segmentation quality.