YOLO-RSD: Enhanced Rotated Ship Detection in SAR Images

Abstract

Rotated ship detection in synthetic aperture radar (SAR) imagery remains challenging due to arbitrary orientations, large-scale variations, and strong sea clutter. This letter presents YOLO-RSD, a lightweight yet effective oriented object detector tailored for maritime surveillance. We first introduce a feature-aligned spatial attention (FASA) backbone that integrates deformable convolution with spatial attention to enhance orientation- and scale-adaptive feature extraction. Then, a FuseDiff neck module is designed to achieve consistent cross-scale semantic fusion via multibranch convolution and context diffusion. Finally, a task-aligned dynamic detection head (TDDH) is constructed using task-specific attention to reconcile classification–localization conflicts. Experiments on SSDD+ and RSDD-SAR datasets demonstrate that YOLO-RSD achieves the state-of-the-art performance with only 2.61 M parameters, outperforming recent YOLO variants and other comparative models in both accuracy and efficiency. The proposed framework offers a general and hardware-friendly paradigm for robust rotated target detection in SAR imagery.