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

Yunze Zhang, Li-Ying Hao, Yifei Li

Abstract: Synthetic aperture radar (SAR) has become increasingly vital in ship detection, owing to the significance of highresolution images it provides. Ship detection from SAR images is a huge challenge, and the main difficulty is the small size and low resolution due to the long observation distance, resulting in high false negative rate. To solve this problem, a new detection model is proposed in this paper called Ship Detection YOLO (SD-YOLO). It improves the detection accuracy of small targets with almost no sacrifice of the algorithm’s real-time performance. Specifically, we improve the C3 module in YOLOv5 called CB-C3 by combining coordinate attention (CA) and Bottleneck. Moreover, in order to improve the ship detection precision and training efficiency, the α-IoU loss function constraint detection bounding box is employed, so that the new detector can more efficiently acquire the position of the ship object. Then, we redesign the neck layer of YOLOv5 with bidirectional feature pyramid network (BiFPN), which can better integrate multi-scale features. Finally, experimental results on multiple public SAR datasets show that the average precision (AP) of SD-YOLO can reach 96.1% in SAR ship detection dataset (SSDD) and 73.2% in large-scale SAR ship detection dataset (LS-SSDD), which is an increase of 2.7% and 7.9% respectively over the previous improvement. Compared with most mainstream algorithms, SD-YOLO has fewer parameters, only 6.79M.

Submission Number: 154