摘要
自新冠疫情爆发以来,人们日常规范佩戴口罩成为了抗击疫情的关键,但人力监管又会增加相关人员被感染的风险,于是提出了一种轻量化口罩检测算法以实现对其佩戴情况的实时监测。以YOLOv5算法为框架,使用改进的EfficientNetV2替换原网络中的主干特征提取网络,减小了网络模型的参数量并提高了精度。提出了使用ECA模块替换EfficientNetV2网络中的SE模块,并使用DIoU-NMS替代原模型中加权NMS的方式,进一步降低了模型的参数量和提高了模型的收敛性,并且提升了对遮挡目标的检测效果。基于网络公开采集到的口罩数据集试验结果表明,所提算法模型参数量下降了44.7%,mAP达到了95.3%,推理速度达到了270.3FPS。提出的算法能够有效的识别人员是否规范佩戴口罩,从而来实现对人员的有效监控。
Since the outbreak of the COVID-19,people's daily standard wearing of masks has become the key to fighting the epidemic.However,human supervision will increase the risk of infection of relevant personnel.Therefore,this paper proposes a lightweight mask detection algorithm to realize the real-time monitoring of their wearing conditions.This paper used the YOLOv5 algorithm as the framework and replaced the backbone feature extraction network in the original network with the improved EfficientNetV2,which reduced the amount of parameters of the network model and improved the accuracy.It was proposed to use the ECA module to replace the SE module in the EfficientNetV2 network,and use DIoU-NMS to replace the weighted NMS in the original model,which further reduced the amount of model parameters,increased the convergence of the model,and improved the detection of occluded targets effect.Based on the test results of the mask data set publicly collected on the Internet,the algorithm model parameters in this paper have dropped by 44.7%,mAP has reached 95.3%,and the inference speed has reached 270.3FPS.The algorithm proposed in this paper can effectively identify whether a person wears a mask in a standard manner,so as to realize effective monitoring of the person.
作者
黄家興
南新元
张文龙
徐明明
HUANG Jia-xing;NAN Xin-yuan;ZHANG Wen-long;XU Ming-ming(College of Electrical Engineering,Xinjiang University,Urumchi Xinjiang 830049,China)
出处
《计算机仿真》
北大核心
2023年第5期541-547,共7页
Computer Simulation
基金
国家自然科学基金项目(61463047)。
