FCOSMask: Fully Convolutional One-Stage Face Mask Wearing Detection Based on MobileNetV3

Abstract

Wearing masks correctly in public is one major self-prevention method against the worldwide Coronavirus disease 2019 (COVID-19). This paper proposes FCOSMask, a fully convolutional one-stage face mask wearing detector based on the lightweight network, for emergency epidemic control and long-term epidemic prevention work. MobileNetV3 is applied as the backbone network to reduce computational overhead. Thus, complex calculation related to anchor boxes is avoided in the anchor-free method, and Complete Intersection over Union (CIoU) loss is selected as the bounding box regression loss function to speed up model convergence. Experiments show that compared to other anchor-based methods, detection speed of FCOSMask is improved around 3 to 4 times on self-established datasets and mean average precision (mAP) achieves 92.4%, which meets the accuracy and real-time requirements of the face mask wearing detection task in most public areas. Finally, a Web-based face mask wearing system is developed that can support public epidemic prevention and control management.