Masked Face Recognition on Limited Training Data

Effectiveness of deep learning-based face recognition systems often relies on having a substantial number of annotated training samples per identity. The addition of new identities to the system necessitates full retraining of the model, making it computationally expensive and time-consuming. Furthermore, the issues become seriously challenging when people wearing masks and glasses to deal with the COVID-19 pandemic. In this paper, we present a novel approach that combines deep learning-based method with data augmentation targeting critical facial landmarks to improve the performance of face recognition in real-life scenarios, such as wearing masks and glasses. Additionally, we embark on an extensive examination to discern the optimal and minimal number of images requisite for the training of efficient face recognition systems for individual subjects. We also construct a new dataset specifically designed for face recognition on limited data on top of the Labeled Faces in the Wild dataset. Experimental results shows that we can achieve high performance in face recognition (i.e., accuracy of 90% and 80% on recognizing people wearing masks and glasses, and 93% on people not wearing any items) by using only 16 training images for each identity. These findings demonstrate the potential of generating augmented data in addressing the challenges of training deep neural network-based masked face recognition with limited training data. Our work provides valuable insights into the possibilities and limitations of face recognition in practical applications, with potential implications for the security enhancement.