基于双重注意力机制的鱼类体表病理识别方法  

A method of identifying fish surface pathology based on dual attention mechanism

作　　者：王一非 袁涛 吴鹏飞 WANG Yifei;YUAN Tao;WU Pengfei(College of Informatics,Huazhong Agricultural University,Wuhan 430070,China;Key Laboratory of Smart Farming for Agricultural Animals,Ministry of Agriculture and Rural Affairs,Wuhan 430070,China)

机构地区：[1]华中农业大学信息学院,武汉430070 [2]农业农村部智慧养殖技术重点实验室,武汉430070

出　　处：《华中农业大学学报》2025年第2期73-82,共10页Journal of Huazhong Agricultural University

基　　金：国家重点研发计划项目(2022YFD1400400)。

摘　　要：为提高鱼类体表病理识别精确度及识别效率,解决当前识别过程中对人工依赖严重且识别准确性低等问题,根据4种发病率高且对鱼类危害大的鱼病构建鱼类体表病理数据集,基于Resnet-18模型进行改进优化,融合空间注意力和SE通道注意力双重注意力机制构建出高识别精度的DBA_Resnet-18模型,并基于该模型开发了鱼病实时智能识别可视化系统。改进后模型将SE通道注意力模块添加在网络中部,在网络尾部引入了空间注意力机制。测试结果显示,DBA_Resnet-18模型对鱼类体表病理分类准确率达到了96.75%,相比于常见的模型Resnet-18、Resnet-34、Resnet-50、Resnet-101、Swin Transformer、VGG-16、VGG-19和AlexNet分别高出1.71、2.12、2.37、2.83、2.51、2.23、2.50和3.53百分点。研究结果表明,本研究提出的模型及鱼病智能识别可视化系统能够对不同鱼类体表病理进行快速、精确的分类识别,实现了鱼病识别系统的智能化,可用于实际环境中诊断鱼类体表病理类型。The dataset of fish surface pathology was constructed based on four types of fish diseases with high rate of incidence and great harm to fish to improve the accuracy and efficiency of identifying fish surface pathology and solve the problems of heavy reliance on manual labor and low accuracy of identification in the process of identification at present.An improved and optimized DBA_Resnet-18 model with high accuracy of identification based on the Resnet-18 model was constructed by integrating spatial attention and SE channel attention dual attention mechanism.A real-time intelligent visualization system for identifying fish diseases was developed based on this model as well.The improved model incorporates SE channel attention module in the middle of the network and introduces spatial attention mechanism at the end of the network.The results of testing showed that the accuracy of the DBA_Resnet-18 model in classifying fish surface pathology reached 96.75%,which was 1.71,2.12,2.37,2.83,2.51,2.23,2.50,and 3.53 percent points higher than that of the commonly used models including Resnet-18,Resnet-34,Resnet-50,Resnet-101,Swin Transformer,VGG-16,VGG-19,and AlexNet,respectively.It is indicated that the proposed model and the developed intelligent visualization system for identifying fish diseases can quickly and accurately classify and identify different fish surface pathologies,realizing the intelligence of the system for identifying fish diseases,which can be used to diagnose the types of fish surface pathology in practical environments.

关 键 词：鱼类体表病理 深度学习 注意力机制 Resnet-18 鱼病智能识别 

分 类 号：S941[农业科学—水产养殖] TP391.41[农业科学—水产科学] TP18[自动化与计算机技术—计算机应用技术]