Face mask detection using deep convolutional neural network and multi-stage image processing

Face mask detection has several applications including real-time surveillance, biometrics, etc. Face mask detection is also useful for surveillance of the public to ensure face mask wearing in public places. Ensuring that people are wearing a face mask is not possible with monitoring staff; instead, automatic systems are a much better choice for face mask detection and monitoring to help manage public behaviour and contribute to restricting the outbreak of COVID-19. Despite the availability of several such systems, the lack of a real image dataset is a big hurdle to validating state-of-the-art face mask detection systems. In addition, using the simulated datasets lack the analysis needed for real-world scenarios. This study builds a new dataset namely RILFD by taking real pictures using a camera and annotating them with two labels (with mask, without mask) which are publicly available for future research. In addition, this study investigates various machine learning models and off-the-shelf deep learning models YOLOv3 and Faster R-CNN for the detection of face masks. The customized CNN models in combination with the 4 steps of image processing are proposed for face mask detection. The proposed approach outperforms other models and proved its robustness with a 97.5% of accuracy score in face mask detection on the RILFD dataset and two publicly available datasets (MAFA and MOXA).