Explainable masked face recognition

The COVID-19 epidemic has made us all to understand that using face masks is one of the best ways to save ourselves from infections. Face recognition techniques often focus on the essential facial landmarks such as nose, mouth and eyes. Removing masks in airports or other public places for authentication will raise the danger of virus infection, and this will pose a challenge to the existing face recognition systems. This paper presents a novel approach to recognize masked faces that combines cropping the unmasked half of the face (the upper part of a face) image using a vision transformer model. Various scenarios are investigated in this paper. The model is trained using the images of the upper half of the face and tested on the same, full-face images with a mask and full face images without a mask. The experiments on the standard datasets, namely RMFRD, MLFW and masked CASIA-WebFace show that the proposed approach improves masked face recognition significantly compared to the existing methods. The explainability of the proposed model is demonstrated using a class activation map.