Contrastive Loss On Masked Face Verification

摘要

The outbreak of COVID-19 has encouraged people to wear their masks more frequently than ever. However, the absence of much facial information from the masked face will cause failures in many current face recognition and verification functions to recognize the individual’s identity. To tackle this problem precisely, our group builds a modified SimCLR model with the contrastive loss that is able to extract similarity features from individuals regardless of whether a mask is worn. From our experiments, we find out that our usage of contrastive loss leads to a large improvement in the testing verification accuracy compared to a baseline model with the commonly used MSE loss.