Boosting Fairness for Masked Face Recognition

Face recognition achieved excellent performance in recent years. However, its potential for unfairness is raising alarm. For example, the recognition rate for the special group of East Asian is quite low. Many efforts have spent to improve the fairness of face recognition. During the COVID-19 pandemic, masked face recognition is becoming a hot topic but brings new challenging for fair face recognition. For example, the mouth and nose are important to recognizing faces of Asian groups. Masks would further reduce the recognition rate of Asian faces. To this end, this paper proposes a fair masked face recognition system. First, an appropriate masking method is used to generate masked faces. Men, a data re-sampling approach is employed to balance the data distribution and reduce the bias based on the analysis of training data. Moreover, we propose an asymmetric-arc-loss which is a combination of arc-face loss and circle-loss, it is useful for increasing recognition rate and reducing bias. Integrating these techniques, this paper obtained fairer and better face recognition results on masked faces.

Automated summary

Face recognition has achieved excellent performance in recent years, but concerns over its unfairness have been raised. This paper proposes a fair masked face recognition system, utilizing appropriate masking methods, data re-sampling techniques, and an asymmetric-arc-loss approach to increase recognition rate and reduce bias in masked faces.