电子测量技术ELECTRONICMEASUREMENTTECHNOLOGY第45卷第23期2022年12月DOI:10.19651/j.cnki.emt.2210035基于改进YOLOv5的复杂场景多目标检测*强栋王占刚(北京信息科技大学信息与通信工程学院北京100101)摘要:针对多目标图像检测环境复杂、目标物位置数据冗余且长宽高数据大小不一的问题,利用神经网络算法可以有效提高不同类目标物并行检测的准确度和稳定性,提出一种基于改进YOLOv5网络的多目标检测方法。首先依据不同目标物的空间尺度大小,改进模型的特征融合方法,添加多尺度特征检测层以减小多目标检测时的误差,同时增加自适应特征增强模块(adaptivefeatureadjustment),降低网络的误检率与漏检率;然后使用K-means++算法估计候选框,获得更优的框参数;最后在损失函数中使用EIOU(efficientIOUloss)做优化。实验表明:改进后的方法mAP(meanaverageprecision)达到76.48%,相比经典YOLOv5网络提升了3.2%,小尺寸目标物检测准确度均值增加6.3%。改进方法网络延续YOLOv5网络的轻量高效,对于多尺度目标物检测获得更优的检测精度,能够实现更准确的实时多目标检测。关键词:神经网络;多目标检测;YOLOv5;自适应特征增强;损失函数优化中图分类号:TP391.4文献标识码:A国家标准学科分类代码:510.4050ImprovedYOLOv5complexscenemulti-targetdetectionQiangDongWangZhangang(SchoolofInformationandCommunicationEngineering,BeijingInformationScienceandTechnologyUniversity,Beijing100101,China)Abstract:Aimingattheproblemsofcomplexmulti-targetimagedetectionscenesandredundanttargetpositiondatawithdifferentlength,widthandheight,theneuralnetworkalgorithmcaneffectivelyimprovetheaccuracyandstabilityofparalleldetectionofdifferenttypesoftargets.Amulti-targetdetectionmethodbasedontheimprovedYOLOv5networkisproposed.First,accordingtothespatialscaleofdifferentobjects,thefeaturefusionmethodofthemodelisimproved,andamulti-scalefeaturedetectionlayerisaddedtoreducetheerrorofmulti-targetdetection.Atthesametime,AdaptiveFeatureAdjustmentmoduleisaddedtoreducethefalsedetectionrateandmisseddetectionrateofthenetwork;thenK-means++algorithmisusedtoestimatethecandidateframetoobtainbetterframeparameters;finally,EfficientIOULossisusedinthelossfunctionforoptimization.Experimentsshowthatthemeanaverageprecisionoftheimp...
