Author Login Editor-in-Chief Peer Review Editor Work Office Work

Computer Engineering ›› 2020, Vol. 46 ›› Issue (11): 12-22. doi: 10.19678/j.issn.1000-3428.0058802

• Hot Topics and Reviews • Previous Articles     Next Articles

Mask Wearing Detection Algorithm Based on Improved YOLOv3 in Complex Scenes

WANG Yihao1, DING Hongwei1, LI Bo1, YANG Zhijun1,2, YANG Jundong1   

  1. 1. School of Information Science and Engineering, Yunnan University, Kunming 650500, China;
    2. Institute of Science and Education, Education Department of Yunnan Province, Kunming 650223, China
  • Received:2020-07-01 Revised:2020-08-22 Published:2020-07-29

复杂场景下基于改进YOLOv3的口罩佩戴检测算法

王艺皓1, 丁洪伟1, 李波1, 杨志军1,2, 杨俊东1   

  1. 1. 云南大学 信息学院, 昆明 650500;
    2. 云南省教育厅科学教育研究院, 昆明 650223
  • 作者简介:王艺皓(1995-),男,硕士研究生,主研方向为深度学习、计算机视觉;丁洪伟,教授、博士、博士生导师;李波,教授、博士;杨志军,研究员、博士;杨俊东,讲师、博士。
  • 基金资助:
    国家自然科学基金(61461053,61461054)。

Abstract: The transmission of COVID-19 virus through respiratory droplets can be effectively prevented by correct mask wearing.However,complex factors in natural scenes including occlusion,crowds,and small-scale targets frequently affect the detection of mask wearing.To solve the problem,this paper proposes a YOLOv3-based mask wearing detection algorithm for complex scenes.The DarkNet53 backbone network is improved based on the cross-stage partial network,which reduces the calculation consumption and increases the training speed.Then an improved spatial pyramid pooling structure is introduced into YOLOv3,and the top-down and bottom-up feature fusion strategies are used to optimize the multi-scale prediction network,so as to realize feature enhancement.In addition,CIoU is selected as the loss function.The distance between the centers of the target and the detection frame,their overlap ratio,and aspect ratio are considered.The experimental results show that compared with the YOLOv3 algorithm,the proposed algorithm improves the detection accuracy of human faces by 7.3% and that of mask wearing by 14.9%,and the detection speed is improved by 6FPS on average.

Key words: YOLOv3 algorithm, mask wearing detection, cross-stage partial network, spatial pyramid pooling, feature fusion, loss function

摘要: 新型冠状病毒可以通过呼吸道飞沫等方式传播,正确佩戴口罩可以有效防止病毒传染,但是自然场景中通常存在遮挡、密集人群和小尺度目标等复杂因素,对人脸佩戴口罩的检测效果产生影响。针对该问题,在YOLOv3算法的基础上,提出复杂场景下的口罩佩戴检测算法。结合跨阶段局部网络对DarkNet53骨干网络进行改进,以降低计算消耗并提高训练速度。在YOLOv3算法中引入改进的空间金字塔池化结构,通过自上而下和自下而上的特征融合策略优化多尺度预测网络,从而实现特征增强。选取CIoU作为损失函数,考虑目标与检测框之间的中心点距离、重叠率以及长宽比信息。实验结果表明,与YOLOv3算法相比,该算法在人脸目标和人脸佩戴口罩目标上的检测精度分别提高7.3%和14.9%,检测速度平均提高6FPS。

关键词: YOLOv3算法, 口罩佩戴检测, 跨阶段局部网络, 空间金字塔池化, 特征融合, 损失函数

CLC Number: