针对小目标的YOLOv5安全帽检测算法  

YOLOv5 Helmet Detection Algorithm for Small Targets

作　　者：李达[1] 刘辉[1] LI Da;LIU Hui(School of Physical and Electronics,Hunan Normal University,Changsha 410081,China)

机构地区：[1]湖南师范大学物理电子科学学院,湖南长沙410081

出　　处：《现代信息科技》2023年第9期9-13,共5页Modern Information Technology

摘　　要：针对当前YOLOv5难以检测小目标、目标识别效果差等问题,提出了一种基于YOLOv5的改进模型。针对开源数据集小目标样本数量不足的问题,重新构建安全帽数据集,扩充小目标数量。引入轻量化的通道注意力ECA模块,提高模型对安全帽的识别能力。将边界框损失函数替换为SIoU加速模型收敛。最后改进Neck部分的特征融合方式,并增加一个小目标检测层。改进算法在自建安全帽数据集上mAP@0.5、mAP@0.5:0.95相较于YOLOv5s分别提高2.6%、1.7%。Aiming at the problems of small target detection difficulty and poor target recognition effect of YOLOv5 at present,an improved model based on YOLOv5 is proposed.In view of the problem of shortage of small target samples in open source data set,the safety helmet data set is rebuilt to expand the number of small targets.A lightweight ECA module is introduced to improve the identification ability of the model to the helmet.The bounding box loss function is replaced by SIoU to accelerate the convergence of the model.Finally,the feature fusion mode of Neck part is improved,and a small target detection layer is added.Compared with YOLOv5s,the improved algorithm on the self-built safety helmet data set mAP@0.5 and mAP@0.5:0.95 respectively increases 2.6%and 1.7%.

关 键 词：安全帽佩戴检测 YOLOv5 ECA注意力 边界框损失函数 小目标检测 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] TP18[自动化与计算机技术—计算机科学与技术]