基于双通道分层协同的CEH-YOLOv8鱼体病害检测方法  

A method of detecting fish diseases with CEH-YOLOv8 based on dual-channel and hierarchical synergism

作　　者：荣弘扬 汤永华[1] 林森 张志鹏 王腾川 刘兴通 RONG Hongyang;TANG Yonghua;LIN Sen;ZHANG Zhipeng;WANG Tengchuan;LIU Xingtong(College of Information Science and Engineering/Liaoning Province Key Laboratory of Machine Vision,Shenyang Polytechnic University,Shenyang 110870,China;College of Automation and Electrical Engineering,Shenyang University of Technology,Shenyang 110159,China)

机构地区：[1]沈阳工业大学信息科学与工程学院/沈阳工业大学辽宁省机器视觉重点实验室,沈阳110870 [2]沈阳理工大学自动化与电气工程学院,沈阳110159

出　　处：《华中农业大学学报》2025年第2期83-93,共11页Journal of Huazhong Agricultural University

基　　金：辽宁省机器人联合基金项目(20180520022);辽宁省应用基础研究计划项目(2023JH2/101300237)。

摘　　要：针对在鱼体病害检测中病害形状不规则、纹理不清晰以及病斑分散导致难以定位真实病害区域的问题,提出一种基于双通道分层协同的CEH-YOLOv8鱼体病害检测方法。首先,提出一种双通道特征提取网络,增强模型对不规则以及不清晰纹理病斑的提取能力。然后,提出一种高效通道空间注意力机制(ECSA),提升模型对分布式目标的识别能力。同时为强化改进后的主干网络,提出一种分层协同的特征金字塔网络(HBFPN),对主干网络提取出的信息进行分层次特征融合,增强模型的特征表达能力。试验结果显示,CEH-YOLOv8网络对鱼体病害的识别精确率、召回率和mAP分别达到83.2%、72.5%和76.2%,相比于SOTA方法 YOLOv10提升了6.9、11.6和11.9百分点,相比原模型提高4.3、6.9和6百分点,单帧图像推理时间为9.1 ms。以上结果表明,改进后的YOLOv8网络可以精准筛选出带病鱼体,可用于提早发现渔业病害以减少经济损失。A method of detecting fish diseases with CEH-YOLOv8 based on dual-channel and hierarchical synergism was developed to solve the problems of the irregular shapes,unclear textures,and scattered disease spots making it difficult to localize the true lesion areas in the detection of fish diseases.A dual-channel feature extraction network was introduced to enhance the ability of model to extract irregular lesion areas with unclear textures.Then,an efficient channel spatial attention(ECSA) mechanism was proposed to improve the capability of model to recognize distributed targets.A hierarchical and balanced feature pyramid network(HBFPN) for was presented to reinforce the improved backbone network and perform hierarchical feature fusion on the information extracted from the backbone network at different levels to enhance the ability of model to express feature.The results showed that the CEH-YOLOv8 network had an accuracy rate of 83.2%,a recall rate of 72.5%,and a mean average precision(mAP) of 76.2%in detecting fish diseases,respectively.Compared with the state-of-the-art(SOAT)YOLOv10 method and the original model,it increased the accuracy rate,recall rate,and mAP by 6.9,11.6,and 11.9 percent points,and 4.3,6.9,and 6 percent points,respectively.The inference time for a single frame was 9.1 ms.It is indicated that the improved YOLOv8 network can accurately screen fish with diseases and be used for early detection of fishery diseases to reduce economic losses.

关 键 词：鱼体病害检测 YOLOv8 特征提取网络 注意力机制 特征金字塔 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]