Synthetic Aperture Radar (SAR) enables all-weather and day-and-night maritime monitoring, yet accurate ship detection remains challenging due to multiplicative speckle, sea/near-shore clutter, and large intra-class appearance variation that can destabilize multi-scale feature fusion and degrade high-IoU localization. In this work, we present LG-RSD-AttnFPN, a practical enhancement to a standard one-stage PAN–FPN detector that addresses SAR-specific nuisance factors from both representation and supervision perspectives. First, we introduce AttnFPN, a lightweight channel-recalibration module inserted after pyramid fusion to suppress clutter-dominant activations and improve the conditioning of multi-scale features, leading to more stable classification and tighter bounding-box regression in complex coastal scenes. Second, we propose LG-RSD (Local–Global Region Self-Distillation), a training-only regularizer that enforces region-level semantic consistency by aligning RoI-pooled embeddings from dense pyramid features (local student view) with crop-based holistic embeddings computed by a frozen teacher (global view) in the same augmented coordinate system. LG-RSD is fully removed during inference, introducing zero test-time overhead. Extensive experiments on two representative SAR ship detection benchmarks, HRSID and SSDD, under a strict COCO-style protocol demonstrate that our method achieves state-of-the-art performance with consistent gains across AP, AP₇₅, and scale-specific metrics, confirming improved localization tightness and robustness to clutter and speckle.
Jing et al. (Fri,) studied this question.