Boosting Masked Face Recognition with Multi-Task ArcFace

Abstract

In this article, we tackle the recognition of faces wearing surgical masks. Surgical masks have become a necessary piece of daily apparel because of the COVID-19-related worldwide health problem. Modern face recognition models are in trouble because they were not made to function with masked faces. Furthermore, in order to stop the infection from spreading, apps capable of detecting if the individuals are wearing masks are also required. To address these issues, we present an end-to-end approach for training face recognition models based on the ArcFace architecture, including various changes to the backbone and loss computation. We also use data augmentation to generate a masked version of the original dataset and mix them on the fly while training. Without incurring any additional computational costs, we modify the chosen network to output also the likelihood of wearing a mask. Thus, the face recognition loss and the mask-usage loss are merged to create a new function known as Multi-Task ArcFace (MTArcFace). The conducted experiments demonstrate that our method outperforms the baseline model results when faces with masks are considered, while achieving similar metrics on the original dataset. In addition, it obtains a 99.78% of mean accuracy in mask-usage classification.