Enhanced SAR Ship Detection in Coastal Scenes: A Unified Framework with Denoising and Dynamic Anchors

Synthetic Aperture Radar (SAR) imagery plays a pivotal role in maritime surveillance, yet detecting ships in complex coastal scenes remains challenging due to noise interference, diverse target scales, and class imbalance. This paper introduces DN-AnchorNet, an integrated detection framework that couples a dedicated denoising branch with dynamic anchor generation to enhance feature extraction and detection accuracy. By incorporating an adaptive weighted Smooth L1 regression loss, the model effectively addresses sample imbalance, improving robustness. Compared with state-of-the-art rotated detectors—including Faster R-CNN, RoI Transformer, Gliding Vertex, YOLOv8-OBB, and H2RBOX-URC—DN-AnchorNet achieves superior overall performance on both RSDD-SAR and SSDD + datasets. It attains the highest AP (0.699 on RSDD-SAR, 0.610 on SSDD+) and F1-score (0.757, 0.689), while maintaining competitive precision and significantly lowering false positives relative to two-stage detectors. The framework effectively balances detection coverage and false-alarm suppression, demonstrating strong applicability for maritime surveillance in noisy, cluttered scenes. All codes are available at https://github.com/yongqi011210/Dn-anchornet.