基于改进YOLOv4的自然人群口罩佩戴检测方法  

A Method on Mask Wearing Detection of Natural Population Based on Improved YOLOv4

作　　者：薛均晓 武雪程 王世豪[1] 田萌萌[1] 石磊[1] XUE Junxiao;WU Xuecheng;WANG Shihao;TIAN Mengmeng;SHI Lei(School of Cyber Science and Engineering,Zhengzhou University,Zhengzhou 450002,China)

机构地区：[1]郑州大学网络空间安全学院,河南郑州450002

出　　处：《郑州大学学报（工学版）》2022年第4期16-22,共7页Journal of Zhengzhou University（Engineering Science）

基　　金：国家自然科学基金资助项目(62006210);河南省高等学校青年骨干教师培养计划(22020GGJS014)。

摘　　要：针对自然场景下的人群口罩佩戴检测常常会受到口罩样式、颜色,佩戴者肤色以及天气等多种因素的影响,提出在原YOLOv4的基础上引入协调注意力机制,进而提升主干特征提取网络对于浅层次特征图像位置信息的利用进而更好地捕获小物体——口罩,同时能够丰富浅层次特征图像的语义信息和加强远距离依赖关系,更精准地定位和识别目标区域;对YOLOv4的网络结构进行改进以提升整体网络的容量以及深度,进而扩大感受野并提升算法的鲁棒性;引入DIoU-NMS在于缓解目标存在遮拦而被错误抑制的现象,DIoU-NMS从IoU指标及检测框的中心点距离两个方面进行非极大值抑制,使得对于IoU阈值的选取没有那么苛刻。实验结果表明,改进YOLOv4的平均精度均值达到95.81%,相较于原YOLOv4平均精度均值提升了4.62%。改进后的YOLOv4具有良好的性能,能够在自然场景下准确地完成口罩佩戴检测任务。The mask wearing detection in natural scenes is often affected by various factors such as the style and color of the mask,the skin color of the wearer,and the weather.In this study,based on the original YOLOv4,the coordinate attention mechanism was introduced to improve the utilization of the backbone network for spatial information of shallow feature maps and better capture small objects-masks.At the same time,it could enrich the semantic information of shallow feature maps and strengthen the long-distance dependencies to more accurately locate and identify object regions.This paper improved the network structure of YOLOv4 to enhance the capacity and depth of the overall network,so as to expand the receptive fields and improved the robustness of the algorithm.The introduction of DIoU-NMS could alleviate the phenomenon that the object was blocked and incorrectly suppressed.DIoU-NMS could perform NMS from the two aspects of IoU and center point distance of bounding boxes,so that the selection of the IoU threshold was not so harsh.The experimental results showed that the average precision of the improved YOLOv4 was 95.81%,which was 4.62%higher than the average precision of the original YOLOv4.The improved YOLOv4 had exciting performance and could complete the task of comprehensive and accurate mask wearing detection in natural scenarios.

关 键 词：深度学习 口罩佩戴检测 YOLOv4 协调注意力机制 神经网络 

分 类 号：TP39[自动化与计算机技术—计算机应用技术]