DS-YOLO: A SAR Ship Detection Model for Dense Small Targets

Detecting dense small targets in Synthetic Aperture Radar (SAR) images has always been a challenge in ship target detection.To address this issue, this paper proposes a ship target detection model for SAR images, named DS-YOLO, which is based on the YOLO11 network architecture.The model introduces Space-to-Depth Convolution (SPDConv) module to enhance the detection capability of small targets.Additionally, a new module, Cross Stage Partial-Partial Pyramid Attention (CSP-PPA), is incorporated to improve the model's ability to extract features at multiple scales and suppress confusing backgrounds.The loss function is optimized using a bounding box loss based on Adaptive Weighted Normalized Wasserstein Distance (AWNWD), enhancing the model's adaptability to images of varying quality.Finally, experiments were conducted on the standard datasets HRSID and SAR-Ships dataset to validate the robustness and reliability of the DS-YOLO model.The experimental results show that, compared to YOLO11n, DS-YOLO achieved an mAP0.5:0.95 of 68.6% on the SAR-Ships dataset and 69.9% on the HRSID, representing improvements of 1.6% and 0.8%, respectively.Additionally, on these small-target datasets, DS-YOLO achieved an mAP0.5:0.95 of 50.8% and 60.4%, representing improvements of 4.2% and 1.2%, respectively, demonstrating higher detection accuracy.