基于YOLOv5轻量化网络的人脸口罩识别方法研究

Research on face mask recognition based on YOLOv5 lightweight network

目的:为解决人脸口罩识别中边缘和移动端设备存储与计算资源受限的问题,提出一种基于YOLOv5轻量化网络的人脸口罩识别方法。方法:选取由主干网络(Backbone)、颈部模块(Neck)和头部模块(Head)组成的YOLOv5模型作为基础框架。首先,使用ShuffleNetv2轻量化网络对Backbone部分进行修改替换;其次,在Neck部分引入Ghost模块和C3_S模块;最后,为提升检测精度,融入卷积块注意力模块(convolutional block attention module,CBAM),形成Shuffle_Yolo_GS_CBAM模型。选用AIZOO数据集训练和验证模型,通过平均精度均值(mean average precision,mAP)、每秒传输帧数(frames per second,FPS)、每秒10亿次的浮点运算数(giga floating-point operations per second,GFLOPS)和参数量评估模型对人脸口罩的识别效果。结果:该模型识别人脸口罩的mAP为89.5%,FPS为158.7帧/s,参数量和GFLOPS分别为2.38 M和4.5 GFLOPS。与YOLOv5s相比,虽然检测精度略有下降,但检测速度提升了39.7%,模型参数量减少了67.3%,模型运算量减少了73.8%。结论:提出的方法在提高检测速度、减少参数量和计算量、保障检测精度方面表现良好,适合部署在边缘和移动端设备上进行人脸口罩识别。

Objective To propose a YOLOv5 lightweight network-based face mask recognition method to solve the problems due to limited storage and computation resources of edge and mobile devices.Methods A YOLOv5 model composed of a backbone network(Backbone),a neck module(Neck)and a head module(Head)was selected as the base framework.Firstly,the Backbone part was modified and replaced using the ShuffleNetv2 lightweight network;secondly,a Ghost module and a C3_S module were introduced in the Neck part;finally,a Shuffle_Yolo_GS_CBAM model was formed by incorporating a convolutional block attention module(CBAM)to improve the detection accuracy.The model was trained and verified with the AIZOO dataset,which was evaluated for face mask recognition by mean average precision(mAP),frames per second(FPS),giga floating-point operations per second(GFLOPS)and parameters.Results The model proposed recognized face masks with the mAP being 89.5%,FPS being 158.7 frames/s,parameters being 2.38 M and and GFLOPS being 4.5 GFLOPS,which had the detection speed enhanced by 39.7%,parameters decreased by 67.3%and operations reduced by73.8%when compared with the YOLOv5s model.Conclusion The method proposed behaves well in increasing detection speed,decreasing parameters and operations and ensuring detection precision,and thus is worthy promoting for face mask recognition on edge and mobile devices.