基于YOLOv5改进模型的丁香鱼围网作业目标检测研究  

Research on target detection of Engraulis japonicus purse seine based on improved YOLOv5 model

作　　者：张佳泽 张胜茂[1] 樊伟[1] 唐峰华[1] 杨胜龙[1] 孙永文 王书献 刘洋 朱文斌[4] ZHANG Jiaze;ZHANG Shengmao;FAN Wei;TANG Fenghua;YANG Shenglong;SUN Yongwen;WANG Shuxian;LIU Yang;ZHU Wenbin(Key Laboratory of Fisheries Remote Sensing,Ministry of Agriculture and Rural Affairs,East China Sea Fisheries Research Institute,Chinese Academy of Fishery Sciences,Shanghai 200090,China;College of Information,Shanghai Ocean University,Shanghai 201306,China;School of Navigation and Naval Architecture,Dalian Ocean University,Dalian Liaoning 116023,China;Key Laboratory of Sustainable Utilization of Technology Research for Fishery Resource of Zhejiang Province,Marine Fisheries Research Institute of Zhejiang,Zhoushan Zhejiang 316021,China)

机构地区：[1]中国水产科学研究院东海水产研究所,农业农村部渔业遥感重点实验室,上海200090 [2]上海海洋大学信息学院,上海201306 [3]大连海洋大学航海与船舶工程学院,辽宁大连116023 [4]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

出　　处：《海洋渔业》2023年第5期618-630,共13页Marine Fisheries

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);国家自然科学基金重点项目(61936014)。

摘　　要：为解决目前鳀(Engraulis japonicus)限额捕捞与分类统计不准确的问题,提出一种改进YOLOv5的识别算法。该方法将SENet注意力机制引入到YOLOv5主干网络结构中,通过融合捕捞作业不同时期的目标信息并降低复杂背景的干扰,以提高模型检测精度和实时检测效率。采用实际拍摄的鳀作业视频,将视频转化为图片格式实现前期标注和处理,对获得的5550幅图像按照8∶1∶1划分训练集、验证集和测试集,设置对照实验,将YOLOv5主干网络替换为MobileNetV2,并引进SENet注意力机制,分别通过4种模型进行对比,结果表明,该识别算法获得平均精度均值(mAP)为99.4%、精度为98.9%、召回率为99.1%,相比原模型分别提高了2.5%、3.7%和2.9%。研究结果可以为鳀围网作业的目标识别提供新的思路,同时也为渔获作业统计提供了一种辅助手段。Engraulis japonicus is a small pelagic fish widely distributed in the East China Sea and Bohai Sea.For a long time,China offshore fishery has shown a trend of overfishing,and the fish population structure has experienced problems such as low age,miniaturization and early first sexual maturity.Therefore,China has been constantly improving the specific implementation details of fishing quota system.For a long time,the statistics of fishing boat operations have mainly relied on manual recording,which often leads to omissions and mistakes,resulting in inaccurate results of fishing statistics.In order to solve the current problem of inaccurate quota fishing and classification statistics of Engraulis japonicus,this paper learns about deep learning related algorithms through literature search and comparison,which are mainly divided into One stage and Two stage.Combined with the actual needs,this paper adopts the YOLOv5 algorithm in One stage algorithm to realize the detection and improve it to amelioerate the detection performance.Through SENet module,the target features can be deeply extracted.The experimental data adopts the double-vessel purse seine fishing operation method to fish Engraulis japonicus.This experimental data uses a two-boat purse seine fishing method for Engraulis japonicus,the operating principle is to use two identical fishing boats towing long net wings symmetrically left and right to surround the fish,and force the fish into the net bag.Specific realization process is that the fishing vessel is around the other fishing vessel for seining operations,and fishing catch is put into the fishing basket,after a number of fishing operations,the fishing baskets containing the catch are transported to the processing vessel for processing.This completes an operation process,and the fishing vessel operates on the main route from week to week,and its main fishing time is daytime.First,the Hikvision model definition camera with a resolution of 1920×1080 is deployed on the fishing boat to shoot the actual video

关 键 词：鳀 围网捕捞 YOLOv5 注意力机制 目标检测 

分 类 号：S9-9[农业科学—水产科学]