摘要
在人流密集场所,人工方式检查人员是否佩戴口罩容易产生漏检问题,利用迁移学习思想,提出一种基于深度学习的Faster R-CNN口罩佩戴检测算法(FMD-RCNN)。为加速模型收敛,将ImageNet数据集上的预训练模型权重迁移到口罩佩戴检测任务中,通过添加权重衰减等优化方法缓解训练中出现的过拟合现象。在FDDB(Face Detection Data Set and Benchmark)和MAFA(Masked Faces)数据集中选取图像建立数据集。实验结果表明,FMD-RCNN算法的检测精度满足人流密集场合的检测需求,测试集精度可达89.41%,可为海关、机场等区域的口罩佩戴检测提供技术支持。
In crowded places,manual inspection of wearing mask is prone to missed inspection.A face mask detection using Faster R-CNN(FMD-RCNN)algorithm based on deep learning is proposed with utilizing the idea of transfer learning.In order to accelerate the convergence of the model,the weight of the pre-training model on ImageNet dataset is transferred to the mask wearing detection task,and the over fitting phenomenon in training is alleviated by adding weight attenuation and other optimization methods.In FDDB(Face Detection Data Set and Benchmark)and MAFA(Masked Faces)data sets,the experimental results show that the proposed model FMD-RCNN can meet the detection requirements of crowded occasions,and the accuracy of the test set can reach 89.41%.Therefore,the model can provide technical support for mask wearing detection in customs,airports and other areas.
作者
任钰
刘全金
黄忠
胡浪涛
刘国明
REN Yu;LIU Quanjin;HUANG Zhong;HU Langtao;LIU Guoming(School of Electronic Engineeering and Intelligent Manufacturing,Anqing Normal University,Anqing 246133,China)
出处
《安庆师范大学学报（自然科学版）》
2021年第4期25-30,共6页
Journal of Anqing Normal University(Natural Science Edition)
作者简介
任钰(1995-),男,江苏宿迁人,安庆师范大学电子工程与智能制造学院硕士研究生,主要研究方向为数字图像处理。E-mail:496889996@qq.com;通信作者:刘全金(1971-),安徽寿县人,安庆师范大学电子工程与智能制造学院教授,主要研究方向为机器学习、图像处理和无线通信优化。E-mail:liuquanjin@aqnu.edu.cn。
