# CoReFace: Sample-Guided Contrastive Regularization for Deep Face Recognition

**Authors:** Youzhe Song, Feng Wang  
**arXiv:** [2304.11668](https://arxiv.org/abs/2304.11668)  
**Date:** 2023-04-23  
**Summary page:** https://www.summarizepaper.com/en/paper/2304.11668v1/

## Key points

- Discriminability of feature representation is crucial in open-set face recognition
- Proposed approach called CoReFace applies image-level regularization in feature representation learning
- CoReFace employs sample-guided contrastive learning to regularize training based on image-image relationship
- Augments embeddings to integrate contrastive learning into face recognition without degrading image quality
- Introduces a novel contrastive loss with adaptive margin and supervised contrastive mask for steady loss values and prevention of collision with classification supervision signal
- Addresses semantically repetitive signal problem in contrastive learning through new pair coupling protocols
- Extensive experiments demonstrate efficacy and efficiency of CoReFace approach in face recognition, competitive with state-of-the-art approaches
- Exploration of new pair coupling protocols helps mitigate semantically repetitive signal problem
- CoReFace addresses issues related to identity representation and inconsistency in face recognition
- Incorporation of contrastive regularization improves discriminability in feature representation learning.

## Summary

The discriminability of feature representation plays a crucial role in open-set face recognition. To address this issue, we propose a novel approach called Contrastive Regularization for Face recognition (CoReFace) to apply image-level regularization in feature representation learning. Our method employs sample-guided contrastive learning to directly regularize the training based on the image-image relationship, which is consistent with the evaluation process. Instead of degrading image quality, we augment embeddings to integrate contrastive learning into face recognition. In order to effectively incorporate contrastive learning into face recognition, we introduce a novel contrastive loss that considers an adaptive margin and a supervised contrastive mask. This helps generate steady loss values and prevents collision with the classification supervision signal. Additionally, we identify and address a semantically repetitive signal problem in contrastive learning by exploring new pair coupling protocols. We conducted extensive experiments to evaluate our CoReFace approach, and the results demonstrate its efficacy and efficiency. Our method proves to be highly competitive with state-of-the-art approaches in face recognition. Furthermore, our exploration of new pair coupling protocols helps mitigate the semantically repetitive signal problem. Overall, our proposed CoReFace approach addresses issues related to identity representation and inconsistency in face recognition. By incorporating contrastive regularization, we achieve improved discriminability in feature representation learning.