Masked Face Recognition: Progress, Dataset, and Dataset Generation

Face recognition is one of the oldest and famous research in computer vision and has improved over the years. However, most research of this topic are designed to recognize full human face. Ever since COVID-19 pandemic, most people must wear face mask when going outside. This create new challenges for existing face recognition as it is proven to be ineffective recognizing people wearing mask. This paper focuses on discussing previous work of Masked Face Recognition, masked face recognition dataset that can be used for training, and dataset generation tools to generate new masked face dataset using existing face recognition dataset. Analysis suggests that cropping based approach is popular among researchers to solve masked face recognition combined with other method like triplet loss implemented in ResNet-50.

I. Introduction

Face recognition is one of the most famous research in computer science. This research field has been studied over the years and the accuracy has improved over the years [1]–[3]. However, a new challenge emerges when a pandemic caused by COVID-19 emerges and forces people to wear face masks most of the time. This caused original face recognition models to be ineffective because most of them are trained using uncovered faces. Face masks usually covers half of face. This means the model that trained using non-masked face must recognize faces using half data than usual. This is the challenge that original face recognition model is facing.