复杂场景下基于YOLOv5的口罩佩戴实时检测算法研究

Research on Real-time Mask-Wearing Detection Algorithm Based on YOLOv5 in Complex Scenes

在新型冠状病毒疫情防控常态化要求下,目前的口罩佩戴检测装置受复杂场景下人员数量多、相互间易遮挡以及待检目标尺度小的影响,易出现误检漏检等情况;为解决以上问题,提出一种基于YOLOv5的口罩佩戴检测算法以实现复杂场景下的实时检测;首先对数据集做Mosaic数据增强等处理;再经过Focus处理为后续的特征提取保留更完整的图片下采样信息,然后利用SPP融合多尺度信息实现特征增强,在Neck部分保留空间信息;最后考虑目标框与检测框之间的重叠面积、中心点距离和长宽比选用CIoU损失函数以提高定位精度,并且在训练过程中对学习率采用动态调整策略;实验结果表明,改进后算法的平均精度均值可达到99.6%。

Under the requirements of the normalization of the prevention and control of the covid-19 pandemic,the current mask-wearing detection device is affected by numerous factors,such as a large number of people in complex scenes,the easy obstruction of the gathering crowd,and the small size of the inspection target,which are prone to false detections and missing inspections.To solve the above problems,a mask-wearing detection algorithm based on YOLOv5 is proposed to realize real-time detection in complex scenes.Firstly,do Mosaic data enhancement and other processes on the data set.Then,apply the Focus process to remain more complete down sampling information of the images for subsequent feature extraction.Later,Feature enhancement using SPP fusion of multi-scale information and retain spatial information within Neck.Finally,considering the overlapping area,center point distance and aspect ratio between target frame and detection frame,CIoU Loss function is selected s to improve the positioning accuracy.And during the training process,a dynamic adjustment strategy is adopted for the learning rate.Experimental results show that the average accuracy of the improved algorithm reaches 99.6%.