改进YOLO v3的面部口罩佩戴检测算法

Face Mask Wearing Detection Algorithm Based on Improved YOLO v3

2020年1月中国爆发新冠肺炎病毒,其中是否正确佩戴口罩对防疫效果有着至关重要的作用。为了降低行人因侥幸心理不佩戴或错误佩戴口罩而引起的病毒传播风险,提出了一种基于改进YOLO v3算法的面部口罩佩戴检测算法。通过改变YOLO v3中的网络结构,建立输出为4倍降采样特征融合目标检测层,提高网络对口罩佩戴问题中微小错误的召回率和检测的准确率;采用多尺度训练策略,提高网络对输入图片尺寸的稳健性。最后,对口罩佩戴检测中可能出现的干扰因素进行研究。实验结果表明,在面部口罩佩戴问题的检测中,改进型YOLO v3对是否佩戴口罩检测的mAP达94.1%,对是否正确佩戴口罩检测的mAP达90.4%,相比于YOLO v3网络,改进后网络检测性能均有较大提升。

China witnessed the outbreak of the COVID-19 in January 2020,and whether to wear masks correctly plays an important role in effect of epidemic prevention. In order to reduce the risk of virus transmission caused by pedestrians without wearing masks or wearing masks by mistake,a face mask wearing detection algorithm based on improved Yolo v3 algorithm is proposed in this paper. By changing the network structure in Yolo v3,a feature fusion target detection layer is established whose feature map is down-sampled by 4× and it effectively improves the recall rate and detection accuracy of micro errors in mask wearing. In order to enhance the robustness of the network,the strategy of Multi-scale training is used. Finally,the possible interference factors in mask wearing detection are studied. The experimental results show that in the face mask wearing problem detection,the mAP(Mean Average Precision)of improved YOLO v3 on detecting whether to wear mask reaches 94.1% and detecting whether to wear mask correctly reaches 90.4%. Compared with the original Yolo v3 network,the performance of network detection has been greatly improved after the improvement.