An Improved YOLOX Based Face Mask Detection Method in Complex Scenes

Proper use of masks can effectively reduce the risk of spreading the epidemic. In the task of detecting whether people wear masks, it often faces the problems of low detection rate of long-distance and close-range personnel wearing masks, high false detection rate when facing sideways, and complex and changeable real-time detection scenes, which lead to difficulties positioning. Therefore, this paper proposes a face mask detection algorithm based on the YOLOX algorithm. A Coordinate Attention (CA) mechanism is embedded in the model to enhance the network’s learning of mask target details. an Adaptively Spatial Feature Fusion (ASFF) module is also added to the model to make full use of the information of masks at the same position in feature layers of different sizes. In addition, Varifocal Loss and EIoU Loss are used to replace the original confidence loss and localization loss, which improves the ability to select the optimal bounding box and target localization. The experimental results show that the improved algorithm has good accuracy and practicability for the mask detection task, and can better meet the actual needs of the current epidemic prevention and control.