基于改进YOLOv5的船舶目标检测算法-《集美大学学报（自然科学版）》

文章信息/Info

Author(s):: SONG Ce; YIN Yong; WANG Peng; (Key Laboratory of Marine Simulation & Control for Transportation Industry, Dalian Maritime University,Dalian 116026,China)

Keywords:: ship target detection; video sencing; convolutional neural network; YOLO; bounding box selection

摘要:: 针对船载视频感知系统存在的船舶目标检测精度不高和检测速度慢的问题，提出一种改进的YOLOv5船舶目标检测方法。利用SENet以及Confluence边界框选取抑制技术对Backbone和边界框的选取方法进行改进，以提高目标检测精度。通过SeaShips公开数据集对改进的算法进行训练及测试。结果显示：算法的召回率为98.3%；精确率可达88.5%；检测速度达到0.019 s/image。表明改进算法具有较高的检测精度，检测效率可以达到实时。

Abstract:: Aiming at the problems of low detection accuracy and slow detection speed in shipborne video sensing systems, an improved YOLOv5 ship target detection method is proposed in this paper. In this method, SENet and Confluence bounding box selection suppression technology is used to improve the Backbone and bounding box selection method, so as to improve the target detection accuracy. By using Seaships public datasets to train and test the improved algorithm, the recall rate of the proposed algorithm is 98.3%, the precision is up to 88.5%, and the detection speed is up to 0.019 s/image. Experimental results show that the improved algorithm proposed in this paper has high detection accuracy and real-time detection efficiency.