复杂场景下实时人脸口罩检测研究

Resarch on real time face mask detection in complex scenes

针对AIZOO开源人脸口罩检测算法FaceMaskDetection存在较严重的人脸口罩分类精度低的缺陷,本文设计了高精度轻量级人脸口罩分类模型,提出快速特征提取模块FastBlock和基于多层级特征融合的轻量级人脸口罩分类网络(Light MaskNet)。FastBlock减少深度可分离(depthwise,DW)卷积和1×1卷积中间张量的通道数量,进一步降低计算成本,从而提高了特征提取速度。不同层级之间的特征融合可以增大模型的广度,提高模型的鲁棒性。实验结果表明,该人脸口罩分类模型精度可达98.852%,中央处理器(central processing unit,CPU)推理时间仅为9.8 ms,图形处理器(graphics processing unit,GPU)可实现亚毫秒级运算,仅牺牲少量计算资源就能弥补FaceMaskDetection精度低的缺陷,可很好地满足计算资源有限的边缘设备、移动端等的应用需求。

Aiming at the serious defect of low classification accuracy in the open source facial mask detection algorithm FaceMaskDetection of AIZOO,a high-precision and lightweight face mask classification model is designed in this paper,proposing a fast feature extraction module named FastBlock and a lightweight face mask classification network based on multi-level feature fusion named Light MaskNet.FastBlock reduces the number of channels between DW convolution and 1×1 convolution tensors,further reducing computational cost and improving feature extraction speed.The experimental results show that the accuracy of the face mask classification model can reach 98.852%,the CPU reasoning time is only 9.8 ms and GPU reasoning time is sub-millisecond.It can make up for the low accuracy of FaceMaskDetection,which can well meet the application needs of edge devices and mobile terminals by consuming a small amount of computing resources.