Research On Classification And Target Detection Using SAR Images

With the development of SAR (Synthetic Aperture Radar) technology, the data collection capability of SAR is growing rapidly, it is difficult for visual interpretation to meet the rapid growth of SAR data. Automated or semi-automated imagery exploitation for SAR images through computer and pattern recognition technology can tremendously improve efficiency of data processing, and has a promising future and application value in not only military but also civil field. SAR image classification, target detection and segmentation are the important parts of automated or semi-automated imagery exploitation, and have become international focus of research. Some research work is done about these research areas in this thesis, and the main contribution of this dissertation includes:(1) Texture analysis is an important part of the SAR image analysis, the comprehensive cognition to texture will improve the understanding of SAR images and target recognition. Methods of texture feature extraction based on GLCM (Grey Level Co-occurrence Matrix) and classification based on SVM (Support Vector Machine) is studied in this thesis and feature extraction method of class feature base based on GLCM is proposed, in which the number of feature parameter is consistent with the types of ground features, and the progress of classification become more simple.(2) Because MRF (Markov Random Field) can characterize the space relation, and is supported by optimization algorithm, it plays more and more important role in SAR image processing. Target detection method based on MRF is studied in this thesis, and an algorithm of target detection with increasing structure energy parameter based on MRF for SAR images was proposed, according to MAP (Maximum A Posterior) criterion, the target detection result is obtained by applying SA (Simulated Annealing) algorithm to maximize the posterior distribution.(3) Because of characteristic of MRA(Multi-Resolution Analysis) of wavelettransform and self-learning and self-organization ability of neural network, they have wide application in image processing, some research on MRA for SAR image is done in this thesis, according to retina receptive field ganglion cell mathematical model proposed by Rodieck, a feature extraction method of target low frequency wavelet tree by nonlinear sampling is proposed, and then PCA(Principal Component Analysis)is adopted to reduce the dimension of low frequency wavelet tree, the results of which are used to train LVQ(Learning Vector Quantization) neural network, then applies trained neural network to detect target in SAR image, good results are achieved.(4) The result of SAR image segmentation directly affect the following method and quality of recognition, more than thousand different types of segmentation algorithms is proposed, among of these methods, active contour segmentation has been paid more attention in study and application. In this thesis active contour segmentation is studied, based on SA algorithm, an improved GVF (Gradient Vector Flow) active contour segmentation method is proposed, which make active contour jump from local optimum solution and converge to global optimum by adding random noise to active contour. Because clutter around target and speckle noise in SAR image disturb the convergence result of active contour, a filter method based on multi-resolution analysis is studied, which reduce speckle noise and clutter around target according to their band, this method improve the performance of active contour segmentation algorithm.(5) Polarimetric information of target can improve the precision of classification, classification method of target decomposition theorems based on coherence matrix and fuzzy C-mean based on Wishart distance measure is studied in this thesis, experiment results are acquired, advantages and disadvantages of which is analyzed. Three types of texture primitive are also summarized and generalized in this thesis.