대표 메모리와 주기적 동결 전략을 활용한 신규 클래스 추가를 위한 새로운 기법

초록

With the rapid advancement of deep learning-based face recognition technology, it is now widely applied in various domains such as enterprise access control systems and personal or home devices. However, when new users (i.e., classes) are added, retraining the model with the entire dataset becomes time-consuming, while training with only the new samples may lead to performance degradation. To address these issues, this paper proposes a training method that combines a representative memory, which stores key samples from existing classes, with a periodic freezing strategy that selectively freezes certain model layers during training. Experimental results show that the proposed method achieves up to 11.57% faster convergence to 100% validation accuracy compared to the baseline in environments using both CPU and GPU. In CPU-only environments, the improvement reaches up to 46.7% faster convergence, demonstrating the effectiveness of the approach in reducing training time while maintaining high performance.

키워드