Learning From Natural Images in Few-Shot SAR Target Classification

The intricate imaging attributes of synthetic aperture radar (SAR) present a formidable challenge to the prevailing few-shot target classification. In order to address this issue, we study how to leverage natural images to assist with few-shot SAR learning and propose a model with cross-domain generalization ability, named CDFS-SAR. The core of our approach is a dual-branch network based on supervised contrastive learning. This network comprises a global branch for holistic image understanding and a local branch for capturing fine-grained image details. By exploiting the potential semantic consistency between these branches, we facilitate the learning of background-separated prior knowledge, enabling effective cross-domain generalization to SAR images. To further address the challenge of category confusion in SAR images, we introduce an embedding space reconstruction module. This module utilizes supervised contrastive learning to enhance intraclass compactness and interclass divergence of features. In addition, we employ a Brownian distance covariance metrics module to overcome the limitations of conventional metric distances, ensuring the model acquires more precise semantic representations. A novel joint loss function is formulated to optimize cross-domain generalization and accommodate the unique imaging characteristics of SAR. Experimental results on the MSTAR and OpenSARShip datasets demonstrate the effectiveness of our method and highlight the ability to improve few-shot SAR target classification.