摘要
在疫情防控常态化的形势下,口罩佩戴检测可以说是每个公共场所的必备操作。利用现有的深度学习相关知识进行口罩佩戴检测,能够解放大量的人力物力,具有重大的实用意义。论文针对口罩佩戴检测因为人员密集,容易互相遮挡导致误检、漏检等情况,在YOLOv4算法的基础上做出了改进。首先使用K-means++算法对数据集内的真实框进行尺寸聚类,提高网络的拟合能力;其次在CIOU损失函数的基础上,使用效果更好的alpha-IOU损失函数,优化训练过程;最后使用Soft-NMS算法替换原有的NMS算法,改善检测过程中因预测框相距过近而相互抑制的情况。实验结果表明,该算法在论文的数据集上有着更高的检测精度,可以有效地进行口罩佩戴检测任务。
Under the situation of normalized epidemic prevention and control,mask wearing detection is a necessary operation in every public place.It is of great practical significance to use existing deep learning knowledge to carry out mask wearing detection,which can liberate a lot of manpower and material resources.This paper improves the YOLOv4 algorithm on the basis of the mask wearing detection due to the fact that there are too many people and it is easy to block each other,resulting in false detection and missed detection.Firstly,K-means++algorithm is used to perform size clustering for real boxes in the data set to improve the fitting ability of the network.Secondly,on the basis of CIOU loss function,alpha-IOU loss function with better performance is used to optimize the training process.Finally,Soft-NMS algorithm is used to replace the original NMS algorithm to improve the mutual suppression of the prediction boxes due to the close distance.Experimental results show that the algorithm has higher detection accuracy in the data set of this paper,and can effectively carry out mask wearing detection.
作者
徐东东
李岳阳
罗海驰
XU Dongdong;LI Yueyang;LUO Haichi(Jiangsu Pattern Recognition and Computational Intelligence Engineering Laboratory,Jiangnan University,Wuxi 214000;School of Internet of Things Engineering,Jiangnan University,Wuxi 214000)
出处
《计算机与数字工程》
2024年第8期2289-2293,共5页
Computer & Digital Engineering
