Multiresolution SAR Target Recognition Based on Physical Attention Enhancement and Scale Distillation

The feature representation of synthetic aperture radar (SAR) targets is sensitive to different sensor attributes due to the coherent imaging mode. Influenced by different sensor configurations, the characterizations within resolution cells of SAR targets from multisource are significantly different. Therefore, the integration of multisource remote sensing data across resolutions is of great significance to improve the performance of SAR automatic target recognition (ATR). In this article, we propose a multiresolution SAR ATR method based on physical attention (PA) and scale distillation. First, the PA enhancement module with incoherent entropy (IE) is designed, which assigns relative salient weights to targets from the sensor perspective. In this way, PA is established to characterize the robust properties of SAR targets. Then dual attention is constituted by combining PA and visual attention; they integrate bottom-up robust representations and top-down task-oriented features. Finally, scale knowledge distillation is proposed to transfer multiscale features, thereby complementing discriminative knowledge beyond label supervision for SAR ATR. Extensive experiments on multiresolution datasets of FUSAR-Ship and OpenSARShip validate our method when compared with novel deep learning ATR models, transfer learning, and knowledge distillation methods.