YOLOv3-Slim for Face Mask Recognition

Abstract

The use of face mask is advised by World Health Organization (WHO) for preventing transmission of Coronavirus disease 2019 (COVID-19). It is of great value to solve the multi-task object detection problem of non-wearing mask, wrong way wearing mask and standard wearing mask. In this paper, a network YOLOv3-Slim based on YOLOv3 is implemented. It's faster than YOLOv3. Detection speed increased from 15.67 fps to 16.89 fps. In the mean time, we found the effect of the difference of inner class on the classification ability of the model. The large error of inner class will reduce the accuracy of the model and make the attention mechanism ineffective. So after changing the labels of the third data set, We add ECA module to our network. YOLOv3-Slim is more accurate than YOLOv4 in face mask recognition based on our data set. The mAP increased from 89.45% to 92.50%.