Synthetic aperture radar target classification via joint sparse representation of multi-level dominant scattering images

Abstract

As one of the key steps in synthetic aperture radar (SAR) target classification, feature extraction plays an important role. In this study, the multi-level dominant scattering images (DSIs) are generated based on the original images, which can better convey the discrimination capability. DSI is generated by selecting a certain number of dominant scattering points in the original image, which describes the relative positions and amplitudes of the dominant scattering centers. In addition, the background noises, which are usually with low intensities, can be effectively eliminated with high efficiency. For a test sample, each of multi-level DSIs is classified by sparse representation classification (SRC) at first and a reliability level is calculated based on the reconstruction errors, which reflects its contribution to the correct recognition of the present sample. A prescreening of the multi-level DSIs is performed based on their reliability levels. Only those DSIs with higher reliability levels than a predefined threshold is used for target recognition. Afterwards, the selected multi-level DSIs are jointly classified by joint sparse representation (JSR). JSR is actually a multi-task learning algorithm, which comprehensively considers the individual tasks as well as their inner correlations. Therefore, it can better make use of the multi-level to improve the target classification performance. The moving and stationary target acquisition and recognition (MSTAR) dataset is used for experimental evaluation. The results illustrate that the proposed method could achieve notably high recognition accuracy over 97.5% under the standard operating condition (SOC). Moreover, the robustness of the proposed method to various extended operating conditions (EOCs), e.g., configuration variants, large depression variation, noise corruption and partial occlusion, is also superior over some state-of-the-art SAR target classification methods by comparison.