CSE-YOLOv5: A Lightweight Attention Guided YOLOv5 Network based on EIoU Loss
Yifei Li, Li‐Ying Hao, Huiying Liu, Yunze Zhang · 2023
Ship detection using synthetic aperture radar (SAR) images poses several challenges, primarily due to the small size and low resolution of ships in these images resulting from long observation distances. To solve this problem, we optimize the fifth generation version of you only look once through channel attention, spatial attention and efficient intersection over union (CSE-YOLOv5). The CSE-YOLOv5 improves the detection accuracy of small targets with almost no sacrifice of the real-time performance of the algorithm. Specifically, we design network of YOLOv5 based on convolutional block attention module (CBAM). In order to improve the detection accuracy and training efficiency, a new loss function constraint detection bounding box is employed, so that the detector can learn the position of the ship target more effectively. Experimental results on multiple public SAR datasets show that the average precision (AP) of CSE-YOLOv5 can reach 96% in SAR ship detection dataset (SSDD) and 69.1% in large-scale SAR ship detection dataset dataset (LS-SSDD), which is an increase of 2.6% and 3.8% respectively over the previous improvement. Compare with the current mainstream object detection algorithms, the model has better object performance.