侧扫声呐检测沉船目标的改进YOLOv5法  

An Improved YOLOv5 Method for Shipwreck Target Detection by Side-Scan Sonar Images

作　　者：汤寓麟 边少锋[1] 翟国君[3] 刘敏 张卫东 TANG Yulin;BIAN Shaofeng;ZHAI Guojun;LIU Min;ZHANG Weidong(College of Electrical Engineering,Naval University of Engineering,Wuhan 430033,China;Troops,Hulu Peninsula 125000,China;Naval Institute of Oceangraphic Surveying and Mapping,Tianjin 300061,China;91001 Troops,Beijing 100841,China;31016 Troops,Beijing 100088,China)

机构地区：[1]海军工程大学电气工程学院,湖北武汉430033 [2]92116部队,辽宁葫芦岛125000 [3]海军海洋测绘研究所,天津300061 [4]91001部队,北京100841 [5]31016部队,北京100088

出　　处：《武汉大学学报（信息科学版）》2024年第6期977-985,共9页Geomatics and Information Science of Wuhan University

基　　金：国家自然科学基金(41974005,41971416,42074074)。

摘　　要：基于YOLOv3模型的侧扫声呐沉船目标检测方法存在小目标漏警率高、模型权重大、检测速度未能满足实时性需求等问题,根据数据集特点,提出基于YOLOv5模型的侧扫声呐海底沉船目标检测方法。在YOLOv5模型的基础框架下,构建YOLOv5a、YOLOv5b、YOLOv5c、YOLOv5d、YOLOv5s、YOLOv5m、YOLOv5l和YOLOv5x 8种不同深度和宽度的模型结构进行对比实验,并选择最优的结构,使用GA+K(genetic algorithm and K-means)算法优化检测框,并对损失函数进行改进。实验结果表明,改进的YOLOv5a模型在交并比阈值设置为0.5和0.5~0.95的平均准确率分别较原始模型提高了0.3%和0.6%,较YOLOv3算法分别提高了4.2%和6.1%,检测速度达到426帧/s,提升了近一倍,更加益于实际应用和工程部署。Objectives The side-scan sonar shipwreck detection method based on the YOLOv3 model has the problems of high miss-alarm rate of small targets,heavy model weight,and slow detection speed that fails to meet real-time requirements.Methods This paper introduces the YOLOv5 algorithm and proposes an improved YOLOv5 model according to the characteristics of the side-scan sonar shipwreck dateset.We test YOLOv5a,YOLOv5b,YOLOv5c,YOLOv5d,YOLOv5s,YOLOv5m,YOLOv5l and YOLOv5x under the basic framework of YOLOv5 with eight different depth and width model structures.Then we choose the best structure by using genetic algorithm and K-means algorithm to optimize the detection frame,and to improve the loss function through complete intersection over union.Results The results show that under the different range of intersection over union as 0.5 and 0.5-0.95,the average precisions of the improved YOLOv5a model are increased by about 0.3%and 0.6%than that of the original model,respectively.Compared with the YOLOv3 model,the average precisions of the improved YOLOv5a model are increased by 4.2%and 6.1%,respectively,and the detection speed reaches 426 frames per second which is almost doubled that of YOLOv3.Conclusions The proposed method is more conducive to practical applications and engineering deployment.

关 键 词：侧扫声呐 沉船 目标检测 YOLOv5模型 损失函数 GA+K算法 

分 类 号：P237[天文地球—摄影测量与遥感]