基金项目:西安市智能兵器重点实验室项目(编号:2019220514SYS020CG042)收稿日期:2022-10-16基于改进YOLOv5的无人机小目标检测方法研究*易华辉,宋文治,黄金香,王雨璇,丁瑞(西安工业大学兵器科学与技术学院,西安710021)摘要:针对无人机航拍图像检测存在小目标检测准确率低以及检测模型计算量过大的问题,提出了一种基于改进YOLOv5的无人机小目标检测方法。首先,针对小目标存在漏检的问题,在YOLOv5的特征提取网络中引入了高效通道注意力机制(ECA)模块,提高对小目标的特征提取能力,进而提高小目标检测精度;其次,针对模型计算量大的问题,将模型中的CBL模块进行改进,把其中的普通卷积替换为Ghost卷积,减少模型参数和计算量,以便于在小型嵌入式设备部署;最后为了进一步优化和改进YOLOv5算法,采用加权损失函数,以充分学习图像特征。在DOTA数据集上进行测试,实验结果表明,改进的模型提升了小目标检测效果,其mAp为73.1%,比原算法提高了1.9%,速度达到了92ms,可以准确地完成无人机航拍小目标检测任务,同时也满足实时性要求。关键词:小目标检测;无人机;注意力机制;Ghost卷积;损失函数中图分类号:TP391.4文献标志码:A文章编号:1009-9492(2023)02-0139-06UAVSmallTargetDetectionBasedonImprovedYOLOv5YiHuahui,SongWenzhi,Huangjinxiang,WangYuxuan,DingRui(CollegeofOrdnanceScienceandTechnology,Xi´anTechnologicalUniversity,Xi´an710021,China)Abstract:AimingattheproblemsoflowaccuracyofsmalltargetdetectionandexcessivecalculationofdetectionmodelinUAVaerialimagedetection,aUAVsmalltargetdetectionmethodbasedontheimprovedYOLOv5wasproposed.Firstly,aimingattheproblemofmisseddetectionofsmalltargets,anefficientchannelattentionmechanism(ECA)modulewasintroducedinthefeatureextractionnetworkofYOLOv5toimprovethefeatureextractionabilityofsmalltargets,therebyimprovingthedetectionaccuracyofsmalltargets.Secondly,inviewoftheproblemoflargecalculationofthemodel,theCBLmoduleinthemodelwasimproved,andtheordinaryconvolutioninitwasreplacedwithGhostconvolutiontoreducethemodelparametersandcalculationamounttofacilitatedeploymentinsmallembeddeddevices.Finally,inordertofurtheroptimizeandimprovetheYOLOv5algorithm,theweightedlossfunctionwasusedtofullylearn...