基于深度学习YOLOV5网络模型的金枪鱼延绳钓电子监控系统目标检测应用  

Application of an electronic monitoring system for video target detection in tuna longline fishing based on YOLOV5 deep learning model

作　　者：王书献 张胜茂[2] 朱文斌[3] 孙永文 杨昱皞 隋江华 沈烈 沈介然 WANG Shuxian;ZHANG Shengmao;ZHU Wenbin;SUN Yongwen;YANG Yuhao;SUI Jianghua;SHEN Lie;SHEN Jieran(College of Navigation and Ship Engineering,Dalian Ocean University,Dalian 116023,China;Key Laboratory of Oceanic and Polar Fisheries,Ministry of Agriculture and Rural Affairs,East China Sea Fisheries Research Institute,Chinese Academy of Fishery Sciences,Shanghai 200090,China;Key Laboratory of Sustainable Utilization of Technology Research for Fishery Resource of Zhejiang Province,Marine Fisheries Research Institute of Zhejiang Province,Zhoushan 316021,China;Liancheng Overseas Fishery(Shenzhen)Company Limited,Shenzhen 518035,China)

机构地区：[1]大连海洋大学航海与船舶工程学院,辽宁大连116023 [2]中国水产科学研究院东海水产研究所农业农村部远洋与极地渔业创新重点实验室,上海200090 [3]浙江省海洋水产研究所浙江省海洋渔业资源可持续利用技术研究重点实验室,浙江舟山316021 [4]深圳市联成远洋渔业有限公司,广东深圳518035

出　　处：《大连海洋大学学报》2021年第5期842-850,共9页Journal of Dalian Ocean University

基　　金：浙江省海洋渔业资源可持续利用技术研究重点实验室开放课题(2020KF001);国家重点研发计划(2019YFD0901405,2019YFD0901402);国家自然科学基金重点项目(61936014)。

摘　　要：为评估金枪鱼延绳钓系统运行质量、降低人工成本,以及从金枪鱼延绳钓系统电子监控EMS系统中提取浮球、金枪鱼数量等信息,本文提出一种基于深度学习YOLOV5网络模型的金枪鱼延绳钓电子监控系统浮球及金枪鱼目标检测方法,从HNY722远洋渔船EMS系统视频监控数据中截取包含有目标浮球和金枪鱼的15578帧关键帧,将所有关键帧及其标记文件划分为14178个训练数据及1400个验证数据,基于YOLOV5s、YOLOV5l、YOLOV5m、YOLOV5x等4种YOLOV5神经网络模型,设计分组训练试验对比训练效果。结果表明:参与训练的4种神经网络模型均可完成金枪鱼延绳钓电子监控系统的目标检测任务,但网络模型的选择对广义交并比损失(GIoU loss)、目标检测损失(objectness loss)、准确率(precision)、召回率(recall)、多类别平均精度值(mAP)等参数具有显著性影响(P<0.05),对目标分类损失(classification loss)参数无显著性影响(P>0.05);检测效果表现较好的模型是YOLOV5l和YOLOV5m,二者的mAP@0.5值分别为99.1%和99.2%,召回率分别为98.4%和98.3%,但YOLOV5m网络模型在GIoU损失等表现上劣于YOLOV5l。研究表明,4种网络模型中YOLOV5l模型是最适合应用于金枪鱼延绳钓电子监控系统目标检测的网络模型。In order to evaluate the operation quality of the tuna longline fishing system,reduce labor costs,and extract information such as float and tuna quantity from the electronic monitoring system of the tuna longline fishing system,a method for detecting floating ball and tuna target in tuna longline fishing electronic monitoring system was proposed based on deep learning YOLOV5 network model.A total of 15578 key frames containing target float or tuna were intercepted from the video surveillance data of the HNY722 ocean-going fishing vessel EMS,and divided all key frames and their mark files into 14178 training data and 1400 verification data,based on YOLOV5s,YOLOV5l,YOLOV5m and TOLOV5x deep learning neural network models.The group training tests were designed to compare training effects.The results showed that the four neural network models trained in this article all completed the target detection task of the tuna longline electronic monitoring system.However,the choice of the network model had a highly significant impact on the parameters of GIoU loss,objectness loss,precision,recall,mAP@0.5,mAP@0.5∶0.95(P<0.05),without significant impact on the classification loss parameters(P<0.05).The better detection results were observed in YOLOV5m network models,with mAP@0.5 values of 99.1%in YOLOV5l network and 99.2%in YOLOV5m network,and the recall rates of 98.4%in YOLOV5l network and 98.3%in YOLOV5m network.However,YOLOV5m was inferior to YOLOV5l in performance such as GIoU loss.The finding indicates that YOLOV5l is the most suitable network model for target detection in tuna longline electronic monitoring system among the four network models of YOLOV5s,YOLOV5l,YOLOV5m and YOLOV5x.

关 键 词：金枪鱼 延绳钓 YOLOV5神经网络 视频信息提取 

分 类 号：S932.2[农业科学—渔业资源] TP391[农业科学—水产科学]