A Network for Detecting Facial Features during the COVID-19 Epidemic

ABSTRACT

The new coronavirus can spread through respiratory droplets, and wearing a mask correctly can effectively prevent the virus from spreading. However, the current detection algorithms are based on unobstructed faces, which affects the detection task when wearing a mask. To solve these problems, a facial feature detection algorithm based on Mtcnn+Mobilenet+GDBT in complex scenes is proposed. First, it can detect whether to wear a mask and the fatigue state of the face. Second, it can set different thresholds according to the facial characteristics of different people, and initialize the characteristics of different frames in 5 seconds. The innovation of our paper the self-adaption characteristics for every person, it avoids measuring everyone by one standard, which is of great significance to the popularization of the product. Then train a dataset of masks and feature points containing 708 images. The experimental results show that compared with the traditional detection network, the new network can effectively detect facial features in the context of the epidemic. The loss we adopt is Focal loss. The lowest loss of net is 0.01 nearly. The feature of this paper is that before the outbreak of the new crown epidemic, the relevant model only designed the detection method of whether or not to wear the mask. On the basis of wearing the mask on the face, there is no research on the algorithm for judging the facial movements such as fatigue and drowsiness. This work adopts a two-step method, using the mature face detection network mtcnn, first detects the face, and then sends it into the mobilenet network for classification and monitoring status. This makes the model parameters smaller and the accuracy higher. Even if it is divided into two steps, the lightweight network of mobilenet can run smoothly on terminal devices, especially mobile devices. As for the fatigue detection system, the extraction of fatigue features with occlusions on the face has not been circulated on the network. Therefore, our team's innovation in fatigue detection based on facial occlusions such as masks has realized more dimensional detection, with stronger robustness and smaller limitations. In the future, it can be used in areas with severe influenza virus, factories, and sterile environments. It will be able to make judgments on the mask wearing conditions of on-site personnel through real-time images collected by the camera and remind people to wear masks correctly. During the epidemic, it is equipped with an on-board video system, which automatically records the driver's facial state. If fatigue performance such as frequent blinking is detected, the monitoring system will immediately issue an alarm to ensure safety. © 2021 ACM.