A Study on Improving Face Recognition Accuracy Through Self-Supervised Learning

다국어 초록 (Multilingual Abstract)

In the context of face recognition, traditional methods have limitations when dealing with out-of-distribution data. To address these challenges, our study leverages the potential of unsupervised training within the transformer architecture. We developed an automatic video processing approach and a two-stage training model. This method utilizes both abundant unlabeled data in the wild and high-quality labeled data to enhance the training process, employing self-supervised contrastive loss and supervised classification loss, respectively. Experimental results demonstrate the superiority of our approach in terms of generalization across diverse data distributions and improved accuracy. This study validates the effectiveness of unsupervised training for face recognition and is expected to contribute to advancements in handling out-of-distribution data.