A Mask Wearing Detection Method based on Improved YOLOv5

Abstract

Since the outbreak of the Novel Coronavirus in December 2019, the coronavirus has rapidly spread to all parts of the globe, posing a major threat to the lives and health of people all over the world. The transmission of novel coronavirus mainly includes droplet transmission and contact transmission. In general, the transmission route of the virus can be blocked by taking temperature measurements in public places, wearing masks and keeping spacing, etc. to stop the spread of the epidemic. In order to effectively monitor the situation of people wearing masks so that prevention and control personnel can take corresponding measures in time, this paper proposed an improved mask detection algorithm based on YOLOv5. The main work of this paper is as follows: a mask detection dataset containing more than 8,000 images and detailed annotations is established for model training and testing of mask detection tasks. The images in this dataset contain a variety of real social scenarios and are annotated in detail according to the requirements of mask detection tasks. The FcaNet attention module is added to the backbone network of YOLOv5, which enables the network to learn the weight of each channel independently and enhance the information transmission between features, so as to strengthen the discrimination ability of the network to the foreground and background. In the process of training iteration, images of different sizes are randomly input to enhance the generalization ability of the model. The original image pyramid was changed to feature pyramid as feature enhancement module, and the input feature layers of different scales were used for feature fusion and repeated enhancement, so as to make full use of the details extracted from the network. Combined with CIoU loss function and label smoothing strategy, the positioning loss function and prediction category loss function of boundary box in mask detection task were optimized. The Mosaic data enhancement method and the learning rate cosine attenuation strategy are used to improve the training efficiency and detection accuracy of the model. In the ablation experiment, several control experiments were conducted to verify the effectiveness of the proposed training strategy and the improvement of network structure. Experimental results show that the algorithm improves the average accuracy of mask wearing detection by 3.3%. Meet the accuracy requirements of mask detection in real scenarios.