Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (20): 208-218.DOI: 10.3778/j.issn.1002-8331.2204-0473

• Graphics and Image Processing • Previous Articles     Next Articles

Improved YOLOv4-tiny Epidemic Collaborative Mask Wearing Detection Method

CHENG Haoran, WANG Xintao, LI Junran, GUO Ziyi, LIU Wei   

  1. College of Communication Engineering, Jilin University, Changchun 130012, China
  • Online:2023-10-15 Published:2023-10-15

改进YOLOv4-tiny的疫情协同口罩佩戴检测方法

程浩然,王薪陶,李俊燃,郭子怡,刘维   

  1. 吉林大学 通信工程学院,长春 130012

Abstract: The wearing of masks plays an extremely important role in epidemic control. In view of the problem of poor real-time detection of mask wearing in large-scale crowds and difficult to deploy, an modified YOLOv4-tiny epidemic collaborative mask wearing detection method is put forward. The algorithm is based on YOLOv4-tiny, and replaces the CSP modules with Resblock-D modules, which reduces the complexity of the feature extraction network and improves the detection speed; the introduction of SPP increases the receptive field of the network, so that the network can meet the image input of any size and enhance the robustness of the algorithm; it introduces a two-layer CA attention mechanism to improve the utilization of the algorithm to ensure the detection accuracy. Compared with the traditional algorithm YOLOv4-tiny, the experimental rasults reveals that the proposed network mAP is improved by 0.5 percentage points to 94.0%, and the detection speed is improved by 3.96 FPS. On the basis of making ensure a small promotion in detection accuracy, the detection accuracy is effectively increased, and the overall performance is improved.

Key words: target detection, YOLOv4-tiny, mask testing, coordinate attention(CA), spatial pyramid pooling, Resblock-D

摘要: 口罩的佩戴对于疫情防控起着极其重要的作用,针对大规模人群下口罩佩戴检测实时性欠佳、难以部署的问题,提出了一种改进YOLOv4-tiny的疫情协同口罩佩戴检测方法。该算法以YOLOv4-tiny为基础,用两个Resblock-D模块替代CSP模块,降低特征提取网络复杂度,提升检测速度;引入SPP,增加了网络的感受野,使网络满足任意尺寸的影像输入,并增强算法的鲁棒性;引入两层CA注意力机制,提高算法的利用率以保证检测精确度。通过实验检测结果表明,相较于原始YOLOv4-tiny,所提网络mAP提升了0.5个百分点,达到94.0%,检测速度提升了3.96?FPS。在保证检测速度有少量提升的基础上有效提高了检测速度,综合性能得到提升。

关键词: 目标检测, YOLOv4-tiny, 口罩检测, 注意力机制(CA), 空间金字塔池化, Resblock-D