부산대학교 도서관

Masked face recognition gained importance after the outbreak of COVID-19 when people started wearing facial masks to protect themselves against disease. The performance of existing face recognition systems is declined on the images of persons wearing face masks because the discriminative facial features are hidden under the mask. In addition, face recognition algorithms proposed for handling masked face recognition scenarios exhibit degraded performance in unmasked face recognition settings. Therefore, there is a need for a solution that can perform well in masked-face scenarios and maintains the same performance while recognizing the unmasked faces. In this paper, we propose a Masked Face Unveiling Model (MFUM), which can be added to the backbone of existing facial recognition systems and improves the performance of existing approaches in masked face settings without the need to retrain the current models. The MFUM utilizes the masked face embedding produced by the existing face recognition model backbone and uses attention augmented residual model for recognition. It enhances the similarity of masked face embedding with the unmasked face embedding of the same identity and diminishes the similarity with the unmasked facial embedding of the different identities. We experimented with different face recognition models’ backbone, MFUM alternative architectures, and attention mechanisms on the MFR2 dataset with real masked faces and the LFW dataset with synthetic masks. The results show that the proposed Masked Face Unveiling-Attention Augmented Dense Residual Unit trained with quadruplet loss outperformed not only other MFUM architectures and losses but also exhibited superior performance as compared to other state-of-the-art algorithms. [ABSTRACT FROM AUTHOR]