SD-YOLO: An Attention Mechanism Guided YOLO Network for Ship Detection

Synthetic Aperture Radar (SAR) is extensively used for vessel detection due to its capability to produce high-resolution images. However, detecting vessels in SAR imagery remains challenging because of the small object sizes and reduced resolution caused by long observation distances, often resulting in high miss-detection rates. To address this issue, this paper introduces a novel detection model-ship detection YOLO (SD-YOLO), which improves small object detection accuracy while maintaining real-time performance. Specifically, we enhance the C3 module of YOLOv5 by incorporating Coordinate Attention (CA) and a bottleneck mechanism, forming the CB-C3 module. Additionally, to increase detection precision and training efficiency, we implement the α-IoU loss function, which better constrains detection bounding boxes, enabling the model to locate ships more accurately. We also redesign YOLOv5's neck layer using a Bi-directional Feature Pyramid Network (BiFPN) to optimize multi-scale feature fusion. Experiments on several public SAR datasets demonstrate that SD-YOLO achieves an Average Precision (AP) of 96.1% on the SAR ship detection data-set (SSDD) and 73.2% on the large-scale SAR ship detection data-set (LS-SSDD), representing improvements of 2.7% and 7.9%, respectively. Furthermore, SD-YOLO is more lightweight than other mainstream algorithms, with only 6.79M.