Research On High Resolution SAR Images Target Shadow Inpaintng And Target Recognition

The increasing improvement of Synthetic Aperture Radar (SAR) resolution made its imageinterpretation more necessary than ever. Target recognition is the core of the high-resolution SARimage interpretation and plays an important role in the military and civil fields. Supported by theAviation Science Foundation, we researched on three key issues of SAR image target recognition,namely image preprocessing, feature extraction and target classification. The main content of thisdissertation is shown as follows:(1) A target shadow inpainting method for high-resolution SAR image is proposed. First weanalyse the significance of target shadow inpainting. Then the classical algorithms of imageinpainting are studied and their limitations when applied to the SAR image shadow inpainting areanalysised. In view of the limitations of classical image inpainiting, we proposed an adaptive shadowinpainting based on exemplar’s similarity. Experimental results show that the proposed methodachieves target shadow inpainting in high-resolution SAR image.(2) A target feature extraction algorithm for high-resolution SAR image is proposed. In thisdisseratation, we research on two kinds of feature extraction algorithms: the one based on L2normcriteria and the other based on L1norm criteria. L2-norm-based feature extractions are sensitive tooutliers, while extant L1-norm-based feature extractions either face the curse of dimensionality orhave the problem of the excessive feature dimension. We present a L1-norm-based bilateraltwo-dimensional principal component analysis for feature extraction. Experimental results show thatthe proposed algorithm is robust to outliers and achieves good performance on target recognition.(3) A new sparse algorithm used for high-resolution SAR image target classification is proposed.The framework of SAR image target recognition based on sparse representation is studied, in whichsolving sparse representation is one of the key steps. In this dissertation, various sparse algorithms arestudied. Most of the extant sparse algorithms need either sparsity as a prior or to set the stepsizemanually. In view of this problem, an adaptive threshold backtracking matching pursuit algorithm isproposed. Experimental results show that the proposed algorithm has better performance on signalreconstruction and target recognition.