LDS-YOLO: a lightweight denoising and small-object enhanced network for SAR ship detection

Abstract

Synthetic Aperture Radar (SAR) provides all-weather and high-resolution imaging capabilities, making it an important data source for maritime ship detection. However, coherent speckle noise and complex background clutter can obscure weak target responses, while the limited computing resources of edge platforms impose additional constraints on model complexity. Achieving accurate SAR ship detection with a lightweight architecture therefore remains challenging. To address these issues, we propose LDS-YOLO, a lightweight SAR ship detection framework based on YOLOv8n. LW-C2f reduces redundant computation through efficient feature transformation while preserving discriminative target information. SG-Deno introduces frequency-domain adaptive refinement to suppress noise-dominated components and retain target-related features. For multi-scale fusion, CAFM performs context-aware feature interaction to alleviate semantic inconsistency across feature levels and strengthen small-target representation in cluttered scenes. LSSE-Head further incorporates high-resolution prediction and parameter sharing to improve the detection of small ships without substantially increasing model complexity. Experiments on the HRSID and SSDD datasets show that LDS-YOLO achieves mAP50 values of 92.2% and 97.4%, respectively, with only 2.2 M parameters and 6.0 GFLOPs. These results indicate that LDS-YOLO provides a favorable balance between detection accuracy and computational efficiency, supporting its practical deployment for resource-constrained SAR maritime monitoring.