Research On SAR Image Processing Based On Subspace Analysis

With achieving detailed all-weather imagery of ground, Synthetic Aperture Radar (SAR) is widely used in military filed. Unmanned Aerial Vehicle (UAV) is powerful military equipment for realization of"zero casualty"wars under the condition of high technology. So the combined application of UAV and SAR will be an important tool of realizing"zero casualty"in the future wars.One of the key techniques of UAV is Automatic Target Recognition (ATR), which makes UAV perform tasks without man in it. UAV-borne SAR can provide sustained high-altitude surveillance and reconnaissance at large standoff ranges, and can generate a profusion of data. The demands on proceesing and interpretation of these images dramatically increase beyond the limit of man's ability. So it is necessary to develop automatic image analysis systems in practical use.In this paper, the applications and development situation of SAR are surveyed, and the key technique of SAR applications-SAR image processing is researched deeply. Two key problems of SAR image processing are studied in detail, including SAR image speckle suppression and feature extraction in SAR ATR. Firstly, the approaches for SAR image despeckling are researched. With the wavelet coefficients of logarithmically transformed SAR image analyzed, wavelet transform and Independent Component Analysis (ICA) are combined to develop an effective approach for SAR image speckle suppression. This method employs ICA to process the wavelet coefficients of images, extracting the independent components which represent the uncorrupt images. Also the suppressing performance of several standard wavelet bases is compared. Secondly, some feature extraction methods based on subspace are mainly studied. An image-based Two-Dimensional Principal Component Analysis (2DPCA) is proposed to extract SAR target features. Then these features are applied to train support vector machine (SVM) classifer for SAR ATR. Experimental results with Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset show that 2DPCA is an effective method for feature extraction. It gives higher recognition rate, and is computationally more efficient than PCA.