基于YOLOv5s-GCE模型检测糙皮侧耳黄斑病  

YOLOv5s-GCE Model-Based Detection of Yellow Spot Disease in Pleurotus ostreatus

作　　者：张志勇[1] 武同辉 刘靖宇[2] 王硕 王宸 张燕青[1] 郑维 ZHANG Zhiyong;WU Tonghui;LIU Jingyu;WANG Shuo;WANG Chen;ZHANG Yanqing;ZHENG Wei(College of Agricultural Engineering,Shanxi Agricultural University,Jinzhong 030801,Shanxi,China;College of Food Science and Engineering,Shanxi Agricultural University,Jinzhong 030801,Shanxi,China)

机构地区：[1]山西农业大学农业工程学院,山西晋中030801 [2]山西农业大学食品科学与工程学院,山西晋中030801

出　　处：《食用菌学报》2025年第1期89-101,共13页Acta Edulis Fungi

基　　金：山西省现代农业产业技术体系建设专项(2024CYJSTX09);山西省研究生教育创新项目(2022Y339)。

摘　　要：为准确高效检测糙皮侧耳(Pleurotus ostreatus)黄斑病,构建基于YOLOv5s的黄斑病检测模型YOLOv5s-GCE。该模型在YOLOv5s模型基础上引入轻量化GhostNet结构,将坐标注意力(coordinate attention,CA)模块嵌入到YOLOv5s主干网络中,并利用增强交并比(enhanced intersection over union,EIOU)损失函数替换原YOLOv5s网络的完整交并比(complete intersection over union,CIOU)损失函数,利用自建的黄斑病数据集,对YOLOv5s-GCE模型进行消融和对比实验,并将该模型部署在RK3588S人工智能开发板上进行测试。结果表明:相比于原始YOLOv5s模型,YOLOv5s-GCE模型的平均精度均值(mean average precision,mAP)为92.7%(提高2.7%),复杂度显著降低,参数量、权重大小和浮点运算量(giga floating-pointoperations per second,GFLOPs)分别降低44.7%、43.4%和47.2%;YOLOv5s-GCE模型的整体性能优于SSD、YOLOv7、YOLOv8n和Faster R-CNN典型的目标检测模型。部署在RK3588S开发板上的YOLOv5s-GCE模型检测速度可达每秒30.49帧,mAP值为90.2%,可以满足糙皮侧耳黄斑病实时检测需求,研究结果为后续研发食用菌病害智能检测装置提供参考。For accurate and efficient detection of yellow spot disease in Pleurotus ostreatus,a model named YOLOv5s-GCE was developed based on the YOLOv5s model.YOLOv5s-GCE integrated a lightweight GhostNet structure,embedded a coordinate attention(CA)module into the YOLOv5s backbone,and substituted the original CIOU loss function of YOLOv5s with the enhanced intersection over union(EIOU)loss function.Using a self-built yellow spot disease dataset,ablation and comparison experiments were conducted on YOLOv5s-GCE.Subsequently,the model was deployed on an RK3588S AI development board for validation.The results showed that YOLOv5s-GCE outperformed YOLOv5s in terms of mean average precision(mAP)(92.7%,2.7%increase over the baseline),complexity(significantly reduced),parameter count(decreased by 44.7%),model size(decreased by 43.4%),and computational cost(decreased by 47.2%in giga floating-point operations per second,GFLOPs).The overall performance of YOLOv5s-GCE was superior to other typical object detection models,such as SSD,YOLOv7,YOLOv8n,and Faster R-CNN.The detection speed of YOLOv5s-GCE deployed on RK3588S development board was 30.49 frames per second with an mAP value of 90.2%,which satisfied requirements of real-time detection of P.ostreatus yellow spot disease.The results provided a reference for subsequent development of intelligent devices for detecting pathogenic diseases in edible fungi.

关 键 词：糙皮侧耳 黄斑病 YOLOv5s 目标检测 

分 类 号：S436.46[农业科学—农业昆虫与害虫防治]