基于RA-YOLOv5s的粮仓害虫检测模型  

Granary Pest Detection Model Based on RA-YOLOv5s

作　　者：杜聪 王赟 刘思雨 宋雪桦[1] DU Cong;WANG Yun;LIU Si-yu;SONG Xue-hua(School of Computer Science and Communication Engineering,Jiangsu University,Zhenjiang Jiangsu 212013,China)

机构地区：[1]江苏大学计算机科学与通信工程学院,江苏镇江212013

出　　处：《计算机仿真》2023年第4期486-491,共6页Computer Simulation

基　　金：国家重点研发计划(2017YFC1600804);江苏省自然科学基金(BK20180860)。

摘　　要：针对粮仓害虫体积较小且个别种类害虫外形相似而难以区分的问题,提出一种RA-YOLOv5s(ResNeXt and Attention-YOLOv5s)粮仓害虫检测模型。先在YOLOv5s的CSP模块支路中分别引入空间注意力机制和通道注意力机制。将主干网络CSP模块中的残差单元修改为ResNeXt残差单元,同时对模型进行轻量化处理,去除重复的残差单元。最后修改颈部网络CSP模块结构,使其与主干网络保持一致。实验结果表明,RA-YOLOv5s相较于其它主流目标检测模型具有更高的害虫检测平均正确率,而且模型更加轻量化。Since most granary pests are small and difficult to distinguish individual pests with similar appearances,an RA-YOLOv5s granary pest detection model is proposed.First,the spatial attention mechanism and channel attention mechanism were introduced respectively into different branches of the CSP module of YOLOv5s.Second,the residual unit of the CSP module in the backbone network was improved to the ResNeXt residual unit.In addition,the model structure was lightweighted to remove the repeated residual units in the CSP module.Finally,the CSP module structure of the neck network was modified to keep it consistent with the backbone network.The experimental results show that compared with other mainstream target detection models,RA-YOLOv5s has a higher average accuracy rate of pest detection,and the model is more lightweight.

关 键 词：粮仓害虫 目标检测 注意力机制 分组卷积 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]