Research on Lightweight Ship Target Detection Algorithm Based on Improved YOLOv5

Abstract

Intelligent ship detection is a fundamental problem in ship traffic services, port management, maritime security, and automated fishery management. However, ship monitoring devices are often deployed on embedded devices, which generally have lower performance than large computers, and the effect of neural network models is usually limited. To reduce the algorithm's parameter volume and FLOPs and improve ship detection accuracy, this paper proposes a DCS-YOLO model based on YOLOv5. In the backbone of the model, an improved ShuffleNetv2 network with an attention mechanism is designed. In the head, the feature map with 32-fold down-sampling is removed, and Ghost convolution is used instead of the convolution structure. Experimental results show that compared with the original YOLOv5 algorithm, the improved algorithm increases the 0. 5:0.95m AP by 2.4% and reduces the model size by 70.3%.