YOLO-EZ:一种高效且轻量化的葡萄病害检测模型  

YOLO-EZ:An efficient and lightweight model for grape disease detection

作　　者：董元和 李恩泽 贾炎 庄泳 徐正阳 DONG Yuanhe;LI Enze;JIA Yan;ZHUANG Yong;XU Zhengyang(College of Computer and Information Engineering,Hubei Normal University,Huangshi 435002,China)

机构地区：[1]湖北师范大学计算机与信息工程学院,湖北黄石435002

出　　处：《湖北师范大学学报（自然科学版）》2024年第4期12-20,共9页Journal of Hubei Normal University：Natural Science

基　　金：2024年湖北师范大学“研究生创新科研”立项建设项目(2024Y031);2024年国家级大学生创业实践项目(2024023)。

摘　　要：随着深度学习技术的快速发展,计算机视觉在农业病害检测中展现了巨大的潜力。针对葡萄病害检测问题,提出了一种集成注意力机制和加权双向特征金字塔网络的YOLO-EZ模型。YOLO-EZ模型基于轻量化的MobileViTv3主干网络,通过SCConv注意力卷积增强特征提取,同时应用BiFPN优化特征融合过程,显著提高了对葡萄病害特征的识别精度。在葡萄病害数据集上进行的广泛实验表明,YOLO-EZ在精确率达到92.8%,召回率达到89.9%,mAP50为93.6%,mAP50-95为73.3%,大多优于对比的多个先进模型。同时,YOLO-EZ在保持高性能的同时,实现了模型参数的减少,参数量仅为5.8 MB,有利于在移动端和边缘计算设备上部署,显示了其在实际应用中的可行性和效率。With the rapid development of deep learning technology,computer vision has demonstrated significant potential in agricultural disease detection.To address the issue of grape disease detection,this paper proposes a YOLO-EZ model integrating attention mechanisms and a weighted pyramid network with a bidirectional feature.The YOLO-EZ model,based on the lightweight MobileViTv3 backbone network,enhances feature extraction through SCConv attention convolution and optimizes the feature fusion process by using BiFPN,which significantly improves the recognition accuracy of grape disease features.Extensive experiments on the grape disease dataset show that YOLO-EZ achieves a precision of 92.8%,a recall of 89.9%,an mAP50 of 93.6%,and an mAP50-95 of 73.3%,outperforming several advanced comparative models.Moreover,YOLO-EZ maintains high performance while reducing model parameters and only has a parameter size of 5.8 MB,which makes it suitable for the deployment on mobile and edge computing devices and demonstrate its feasibility and efficiency in practical applications.

关 键 词：计算机视觉 农业病害检测 注意力机制 葡萄病害 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]