Real-time Face Mask Detection: Challenges and Solutions using Adaptive YOLOv3

Abstract

In Real-time face mask detection presents Some of the problems, which arrive while detecting face masks in real-time, are the changes in lighting conditions, the presence of occlusions, and different types of masks people wear. To that end, a reliable detection system is developed with proposed Adaptive YOLOv3 integrated with hybrid atrous convolution and contractive ensemble residual learning. Each Adaptive YOLOv3’s layout allows for fine-tuning of features and real-time analysis; critical in dynamic conditions. Hybrid atrous convolution improves the contextual information by changing the receptive field in order to capture multi-scale features which are important for identification of the masks regardless the conditions. The first module of the residual joint learning network is a whole body model which feeds into the other models; detections are then adjusted and accuracy enhanced in each model. The final refinement of the regions of interest is achieved through the post-processing of the detected bounding boxes through Non-Maximum Suppression (NMS) and by setting the confidence thresholds to optimize both, precision and recall. The proposed system does help to solve the problem with real-time detection very well; as a result, the proposed system offers fast and accurate face mask detection that is necessary for tracking people’s compliance with health requirements.