基于3-2D融和模型的毛虾捕捞渔船行为识别  

Recognition of Acetes chinensis fishing vessel based on 3-2D integration model behavior

作　　者：张佳泽 张胜茂[2] 王书献 杨昱皞 戴阳[2] 熊瑛[4] ZHANG Jiaze;ZHANG Shengmao;WANG Shuxian;YANG Yuhao;DAI Yang;XIONG Ying(College of Information,Shanghai Ocean University,Shanghai 201306,China;East China Sea Fisheries Research Institute,Chinese Academy of Fishery Sciences/Key Laboratory of Fisheries Remote Sensing,Ministry of Agriculture and Rural Affairs,Shanghai 200090,China;School of Navigation and Naval Architecture,Dalian Ocean University,Dalian 116023,China;Jiangsu Marine Fisheries Research Institute,Nantong 226007,China)

机构地区：[1]上海海洋大学信息学院,上海201306 [2]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 [3]大连海洋大学航海与船舶工程学院,辽宁大连116023 [4]江苏省海洋水产研究所,江苏南通226007

出　　处：《南方水产科学》2022年第4期126-135,共10页South China Fisheries Science

基　　金：国家自然科学基金重点项目(61936014);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);江苏省农业综合执法专项(2020-SJ-018);中国水产科学研究院基本科研业务费资助(2020TD82)。

摘　　要：针对中国毛虾(Acetes chinensis)产量逐年锐减问题,中国开始对近海海域实施毛虾限额捕捞措施,采用视频监控技术辅助捕捞管理。提出一种基于3-2D融和模型的毛虾捕捞渔船行为识别方法,为限额捕捞管理提供新的解决方案。通过在毛虾渔船上4个固定位置安装高清摄像设备,并记录捕捞作业全过程,共获取600余个视频监控数据作为初始数据;从初始数据中筛选有效的视频数据,同时对视频数据进行5种行为的划分和标记。为了提高网络训练的效率,对视频数据进行压缩和帧数分割等预处理;最后,通过搭建3-2D融合的卷积神经网络来训练模型,实现渔船行为特征的提取和分类。结果表明,捕捞渔船行为识别方法的分类精度为95.35%,召回率为94.50%,平均精确度为96.60%,模型整体得分达93.32%,平均检测时间为35.46 ms·帧^(−1),可用于毛虾渔船捕捞视频的实时分析。Since the yield of Acetes chinensis has decreased sharply year by year,China has begun to implement quota fishing measures for A.chinensis in offshore waters by using video surveillance technology to assist the fishing management.This paper proposes a method for identifying the behavior of A.chinensis fishing vessels based on the 3-2D fusion model,so as to provide a new solution for quota fishing management.By installing high-definition camera equipment at four fixed positions on the A.chinensis fishing vessel and recording the entire process of fishing operations,we had obtained more than 600 video surveillance data had been as initial data.Secondly,we filtered effective video data from the initial data,and divided and labeled them with five behaviors.In order to improve the efficiency of network training,we preprocessed the video data such as compression and frame number segmentation.Finally,the model was trained by building a 3-2D fusion convolutional neural network to realize the extraction and classification of fishing vessel behavior characteristics.The results show that the classification accuracy of the fishing vessel behavior recognition method was 95.35%;the recall rate was 94.50%;the average accuracy was 96.60%;the overall score of the model could reach 93.32%;and the average detection time was 35.46 ms·frame^(−1).The method can be used for real-time analysis of the fishing video of A.chinensis fishing boats.

关 键 词：中国毛虾 限额捕捞 深度学习 卷积神经网络 

分 类 号：S975[农业科学—捕捞与储运]