Target recognition of synthetic aperture radar images based on matching and similarity evaluation between binary regions

Abstract

This work uses the binary target region for synthetic aperture radar (SAR) automatic target recognition (ATR). Due to the differences of physical sizes and target shapes, the region residuals among the same classes and those between different targets are distributed in different manners. The Euclidean distance transform is then performed on the region residuals to further enhance such differences, which is beneficial for correctly discriminating different targets. Based on the results, a similarity measure is formed according to the distribution characteristics of the region residuals. In addition, the designed similarity measure considers the possible variations of the target region caused by the nuisance conditions like noise corruption, partial occlusion, etc. Owing to its robustness and comprehensiveness, the similarity measure is applied to target recognition by comparing the test sample with different kinds of template classes. Experiments are undertaken on the Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset under standard operating condition (SOC) and some representative extended operating conditions (EOCs), i.e., configuration variants, depression angle variation, noise corruption, resolution variation and partial occlusion. Moreover, the proposed method is examined under reduced training size and possible azimuth estimation errors for a comprehensive evaluation. The experimental results demonstrate the superiority of the proposed method in comparison with several baseline algorithms in SAR ATR.