A Deep Convolutional Neural Network to Limit Virus Spread Using Facial Mask Segmentation

ABSTRACT

Due to the recent COVID-19 outbreak the world has experienced many challenges. Limit and control the virus spread rate is one of them. This letter focuses on limiting the speed of virus spreading by monitoring the use of facial mask in crowded public environments such as tourism places, commercial centres, etc. The proposed method first accurately localizes faces using a state-of-the-art approach and, segments facial mask in a second step. The facial mask segmentation allows to distinguish whether the current subject is wearing a facial mask or not but also if it is properly covering the human face. Indeed, most recent face detection algorithms provide as output a set of facial features such as nose tip and mouth corners. By combining these facial features with facial mask segmentation, the proposed method detects real-time subjects that indirectly encourage virus spread in crowded environments. The proposed facial mask segmentation model is trained with pairs of RGB images and its corresponding alpha image created by extending the publicly available real-world masked face dataset. Further, the proposed model is pruned and optimized using the TensorRt library to be usable for real-world applications.