Research Of Classification And Automatic Target Recognition Using SAR Imagery

Along with the development of SAR (Synthetic Aperture Radar) technologies, the datacollection capability of SAR is growing rapidly. It is difficult for visual interpretation tomeet the rapid growth of SAR data. Automated imagery exploitation for SAR imagesthrough computer and pattern recognition technology can tremendously improve efficiencyof data processing. Classification and automatic target recognition using SAR imagery hasbecome an active research area. Classification and automatic target recognition using SARimagery is studied in this dissertation. Main contributions include:1) There are some existing supervised learning algorithms for SAR terrainclassification. All these algorithms require a large number of manually labeled samples fortraining the classifier accurately. However, the manually labeling is time consuming andrequires experienced human annotators. AdaBoost is a statistical learning algorithm basedon small training samples. In this dissertation, we propose a SAR terrain classificationmethod based on AdaBoost. Experimental results demonstrate that compared withtraditional supervised classification algorithms, effective error reduction can be achieved.2) Theoretical studies of learning have focused almost entirely on learning binaryfunctions, including AdaBoost and SVM. In real-world learning tasks, howerer, there aremany multiclass learning problems, including SAR terrain classification. There are manyways to reduce a multiclass problem to multiple binary classification problems. In thisdissertation, we propose two methods: 1. AdaBoost.ECOC based on AdaBoost and ErrorCorrecting Output Codes; 2. Algorithm based on Naive Bayes and MetaClass.3) Manually labeling is time consuming and requires experienced human annotators.In a SAR terrain classification system, there are often much more unlabeled image dataavailable than the labeled ones. The dissertation addresses this issue by combining the hugeamount of unlabeled data and limited labeled data, namely, semi-supervised learning. Basedon the EM algorithm and Co-Training, we propose Agreement-Training and Combination-Training. Finally, we combine semi-supervised leaming and active learningfor SAR terrain classification. Experimental results demonstrate that our method has betterperformance.4) A SAR automatic target recognition system is proposed, which is based on targetdetection, invariant feature extraction, pose estimation and target classification. First, wepropose a kind of Constant False Alarm Rate (CFAR) target detector to find potentialtargets from scenario. Then Scale Invariant Feature Transform (SIFT) is used to extractkeypoints as rotation and scale invariant feature. Pose estimation method based on mutualinformation is proposed and the feature templates whose pose are close to the estimatedpose are selected. Finally, algorithms based on Hidden Markov Model (HMM) andexemplar-based categorization are used to recognize targets. Experimental results show thatthis method can achieve high recognition probability.