SAR Image Target Segmentation And Feature Extraction

SAR image segmentation,feature extraction and feature analysis play a key role in Synthetic Aperture Radar Automatic Target Recognition(SAR ATR).Because of the high sensitivity of the SAR imaging against target aspect,the target aspect will greatly affect the performance of feature analysis and target recognition.This dissertation focuses on three basic components in SAR ATR including SAR target image segmentation,target aspect estimation,target feature extraction and analysis.Aiming at the problem of target image segmentation,a soft segmentation method is studied and implemented.In this method,the structure prior of the image is described by the continuous MRF model,and then an iterative algorithm which approximates MAP estimation of segmentation image is presented based on E-M algorithm.When the algorithm starts to work,the CFAR detection based segmentation with Rayleigh distribution is selected as the initial value for iteration,which can avoid the local optimum and accelerate the convergence.Based on the structure of range leading edge closed to the radar in the SAR target images and the sparse distribution characteristic of the outliers on the leading edge,a robust algotithm is proposed to detect and eliminate the outliers and estimate the target aspect.Moreover,a new method is presented to eliminate the ambiguity of horizontal and vertical aspect estimation based on the length-width ratio of the target region. Experimental results with the MASTAR data show the validity and robustness of the algorithm.Finally,using the MSTAR dataset,the dissertation discusses feature extraction and furtherly studies feature separablity.Because of the small samples of the SAR training images and the high sensitivity of features to target aspect,a Support Vector Data Description(SVDD) method is applied to describe the local spatial distribution of features in each aspect sector and one kind of feature separability measure is proposed. Moreover,this dissertation validates the feature analysis using the recognition framework based on the feature space trajectory(FST) representation concept,which exploits the nature that the features at different aspect views lie on a low dimension curve in feature space.To overcome the drawbacks of linear FST,the weighted FST and kernel FST are also presented.