A Physics-Guided Reconstruction and Injection Framework With Semantic Consistency Constraints for Oriented SAR Target Detection

Abstract

Synthetic aperture radar (SAR) target detection is pivotal for a wide range of applications, including maritime surveillance, reconnaissance, and strategic security. However, complex scattering mechanisms often render SAR targets discrete and discontinuous, leading to severe missed detections and false alarms. To address this, this article proposes the physics-guided reconstruction and injection framework (PRIF), which introduces the attributed scattering center (ASC) model to characterize these discrete physical scattering properties. The framework comprises the topology-guided semantic consistency matching (TSCM) module, the part-aware label assignment (PALA) strategy, and the physics-guided reconstruction and injection network (PRI-Net). Specifically, the TSCM module ensures the semantic consistency of physical parts by resolving the label permutation problem in clustering, while the PALA strategy effectively mitigates the scattering voids issue by achieving spatial alignment between sampling points and physical entities. The core architecture, PRI-Net, effectively boosts part-level local feature discrimination and significantly improves model interpretability. It achieves this by internalizing physical priors into feature representations through the physics-guided progressive reconstruction branch (PPRB) and performing explicit feature injection via the decoupled spatial attention mechanism of the physics-injected dual-drive detection branch (PDDB). The experimental results on the FAIR-CSAR dataset demonstrate that PRIF achieves an mAP50 of 0.468 and an $F1$ -score of 0.547, significantly outperforming existing state-of-the-art detectors. Furthermore, cross-dataset validation on the RSAR dataset and portability experiments on baseline models demonstrate its robust generalization capabilities and plug-and-play characteristics.