基于全局图推理与改进三维动态卷积的鱼类摄食行为分析  

Fish feeding behavior analysis based on global graph reasoning and improved three-dimensional dynamic convolution

作　　者：丁寅 陈明[1,2] 栗征 薛江浩 DING Yin;CHEN Ming;LI Zheng;XUE Jianghao(College of Information Technology,Shanghai Ocean University,Shanghai 201306,China;Key Laboratory of Fisheries Information,Ministry of Agriculture and Rural Affairs,Shanghai 201306,China;Shanghai Shilin Information Technology Co.,Ltd.,Shanghai 201314,China)

机构地区：[1]上海海洋大学信息学院,上海201306 [2]农业农村部渔业信息重点实验室,上海201306 [3]上海时粼信息科技有限公司,上海201314

出　　处：《江苏农业学报》2024年第10期1863-1874,共12页Jiangsu Journal of Agricultural Sciences

基　　金：广东省重点领域研发计划项目(2021B0202070001)。

摘　　要：本研究提出一种基于时间动作检测的轻量化视频分类网络,旨在解决水产智能化养殖中饵料的投喂不均和水体污染等问题,提高投喂准确性和效率。该网络以ResNet 3D为基础,引入深度可分离卷积模块和三维动态卷积模块,以降低模型规模和参数量;同时采用图卷积全局推理模块和稠密卷积模块构建区域和全局关系,增强网络深层特征的表达,提高网络分类准确率。经试验验证,该模型检测准确率可达96.70%,相较变分自动编码器卷积网络和3D ResNet-GloRe网络,其准确率分别提高7.7个百分点和4.4个百分点;同时,该模型的参数量和计算量也明显降低,分别为1.10 M和3.87 G。研究结果表明,该基于时间动作检测的轻量化视频分类网络可以有效提高水产养殖中饵料的智能化投喂的准确性和效率,减少饵料投喂不均以及水体污染等问题,具有较高的应用价值。In this study,a lightweight video classification network based on temporal action detection was proposed to solve the problems of uneven feeding of bait and water pollution in intelligent aquaculture,and improve the accuracy and efficiency of feeding.Based on ResNet 3D,a deep separable convolution module and a three-dimensional dynamic convolution module were introduced into the network to reduce the model size and parameter quantity.The graph convolution global inference module and DenseBlock module were used to construct the regional and global relationships,which could enhance the expression of deep features of the network and improve the classification accuracy of network.Experimental results showed that the detection accuracy of the model could reach 96.70%,which was 7.7 percentage points and 4.4 percentage points higher than that of the variational autoencoder convolutional network and 3D ResNet-GloRe network,respectively.The number of parameters and calculation amount of the model were also significantly reduced,which were 1.10 M and 3.87 G,respectively.Therefore,the lightweight video classification network based on time motion detection could effectively improve the accuracy and efficiency of intelligent feeding in aquaculture,reduce the problems of uneven feeding of bait and water pollution,and had high application value.

关 键 词：鱼类行为 机器视觉 视频分类 全局图推理 动态卷积 

分 类 号：S951.2[农业科学—水产养殖]