TSMAL: Target-Shadow Mask Assistance Learning Network for SAR Target Recognition

Deep learning-based synthetic aperture radar (SAR) target recognition methods mainly emphasize the amplitude characteristics resulting from backscatter at the target's principal scattering points. Shadows, as critical by-products of SAR imaging, encapsulate vital details regarding the target's structural configuration. In order to effectively utilize the target information and shadow information in SAR images, we propose a novel target-shadow mask assistance learning (TSMAL) network for SAR target recognition. It systematically leverages domain knowledge in SAR images through three key points: data preprocessing, network structure, and multitask loss function. Specifically, the data preprocessing, with the help of the segmentation algorithm, extracts the target mask and shadow mask as domain knowledge in the SAR image to be utilized. Then, the target-shadow mask assistance (TSMA) layer is designed to learn complementary representations within each convolutional layer by exploiting target-shadow information. The TSMA enhances the scatter features related to the target regions as well as the shape features contained in the shadow regions, and suppresses backgrounds. Meanwhile, the multilayer coordinate attention (MCA) is used for multiscale feature fusion. Finally, a multitask learning loss is designed depending on the recognition and feature optimization tasks to guide the network learning. By synergistically employing a TSMAL alongside multitask learning strategies, the network proficiently acquires both target and shadow features. The experiments conducted on the moving and stationary target acquisition and recognition (MSTAR) dataset show that the proposed method can effectively enhance the feature extraction ability in the target and shadow regions, and improve the performance of SAR target recognition.