扩展YOLOv5安全帽多级目标分类检测①金源1,张长鲁21(北京信息科技大学计算机学院,北京100101)2(北京信息科技大学经济管理学院,北京100192)通信作者:金源,E-mail:jinyuan@bistu.edu.cn摘要:YOLO是目前计算机视觉目标检测领域比较重要的算法模型之一.基于现有YOLOv5s模型提出了一种扩展的YOLOv5多级分类目标检测算法模型.首先,对LabelImg标注工具进行功能扩展,使其满足多级分类标签文件构建;其次在YOLOv5s算法基础上修改检测头输出格式,在骨干网络前端引入DenseBlock、Res2Net网络模型核心设计思想,获取丰富的多维度特征信息,增强特征信息的重用性,实现了YOLO多级分类目标检测任务.在开源安全帽数据集上同时以安全帽颜色作为二级分类进行训练验证,平均精度,精确率和召回率分别达到了95.81%、94.90%和92.54%,实验结果验证了YOLOv5多级分类目标检测任务的可行性,并为目标检测及多级分类目标检测任务提供一种新的思路和方法.关键词:目标检测;YOLOv5;多级分类;DenseBlock模块;损失函数;深度学习引用格式:金源,张长鲁.扩展YOLOv5安全帽多级目标分类检测.计算机系统应用,2023,32(2):139–149.http://www.c-s-a.org.cn/1003-3254/8946.htmlExtendedYOLOv5forMulti-levelTargetClassificationDetectionofHelmetJINYuan1,ZHANGChang-Lu21(ComputerSchool,BeijingInformationScienceandTechnologyUniversity,Beijing100101,China)2(SchoolofEconomicsandManagement,BeijingInformationScienceandTechnologyUniversity,Beijing100192,China)Abstract:YOLOisoneofthemostimportantalgorithmmodelsinthetargetdetectionofcomputervision.GiventheexistingYOLOv5smodel,anextendedYOLOv5algorithmmodelformulti-levelclassificationtargetdetectionisproposed.Firstly,thefunctionoftheannotationtoolLabelImgisextendedtoconstructmulti-levelclassificationlabelfiles.Secondly,theoutputformatofthedetectionheadismodifiedonthisbasisoftheYOLOv5salgorithm,andthecoredesignideaoftheDenseBlockandRes2Netnetworkmodelisintroducedinthefrontendofthebackbonenetworktoextractrichmul-ti-dimensionalfeatureinformation,enhancethereusabilityoffeatureinformation,andrealizethetaskofYOLO-basedmulti-levelclassificationtargetdetection.Thehelmetcoloristakenasthesecondaryclassificationfortrainingandverificationontheopensourcehelmetdataset,andthea...
