基于YOLOv5增强模型的口罩佩戴检测方法研究  

Research on Mask Wearing Detection Method Based on YOLOv5 Enhancement Model

作　　者：彭成 张乔虹[1] 唐朝晖 桂卫华[2] PENG Cheng;ZHANG Qiaohong;TANG Zhaohui;GUI Weihua(School of Computer Science,Hunan University of Technology,Zhuzhou,Hunan 412007,China;School of Automation,Central South University,Changsha 410083,China)

机构地区：[1]湖南工业大学计算机学院,湖南株洲412007 [2]中南大学自动化学院,长沙410083

出　　处：《计算机工程》2022年第4期39-49,共11页Computer Engineering

基　　金：国家自然科学基金面上项目(61871432,61771492);湖南省自然科学基金(2020JJ4275,2019JJ6008,2019JJ60054);湖南省大学生研究性学习和创新性实验计划项目(S201911535027)。

摘　　要：人脸口罩佩戴检测是公共场所疫情防控中极为重要的措施,智能、高效地检测口罩佩戴情况对实现疫情防控的自动化和数字化具有重要意义。使用卷积类深度神经网络实现端到端的人脸口罩佩戴检测具有可行性,但卷积类神经网络具有结构复杂、参数量和浮点计算量庞大的特点,从而产生较高的计算开销和内存需求,极大地限制了其在资源有限的终端设备上的应用。为了使人脸口罩佩戴监督功能更易获取,并实现多尺度条件下的模型压缩和加速检测,提出一种基于改进YOLOv5的轻量化增强网络模型。设计参数量和计算量更小的GhostBottleneckCSP和ShuffleConv模块并替换原YOLOv5网络中的C3及部分Conv模块,以降低特征通道融合过程中的计算量并增强特征表达能力。实验结果表明,该模型的识别精度达95%以上,模型在精度近乎无损失的前提下,参数量和计算量分别仅为原YOLOv5网络的34.24%和33.54%,且在GPU和CPU上的运行速度分别提升13.64%和28.25%,降低了模型对内存存储及计算能力的要求,更适用于在资源有限的移动端部署。Face mask wearing detection is a very important measure in epidemic prevention and control in public places.Intelligent and efficient detection of mask wearing is of great significance in realizing the automation and digitization of epidemic prevention and control.Convolution deep neural networks are feasible for end-to-end face mask wearing detection,but convolution deep neural networks have a complex structure,large number of parameters,and floating-point calculation,resulting in high computation overhead and memory requirements,which greatly limit their application in terminal devices with limited resources.To facilitate the face mask wearing supervision function and realize model compression for accelerated detection under multi-scale conditions,a lightweight enhanced network model based on improved YOLOv5 is proposed.Design GhostBottleneckCSP and ShuffleConv modules with fewer parameters and less computation replace C3 and some Conv modules in the original YOLOv5 network,to reduce the computation overload in the process of feature channel fusion and enhance the ability of feature expression.The experimental results show that the recognition accuracy of the model is more than 95%.On the premise of almost no loss of accuracy,the number of parameters and calculation load of the model are only 34.24%and 33.54%of the original YOLOv5 network,respectively,while the running speed on GPU and CPU is increased by 13.64%and 28.25%respectively,thereby reducing the requirements of the model for memory storage and computing power,making it more suitable for deployment on mobile terminals with limited resources.

关 键 词：深度学习 口罩佩戴检测 YOLOv5网络 GhostBottleneckCSP模块 ShuffleConv模块 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]