Regularization Using Noise Samples Identified by the Feature Norm for Face Recognition

Abstract

Face recognition involves comparing two facial images to determine whether they represent the same individual. This task can be applied in a variety of environments, including uncontrolled settings such as surveillance systems, where image quality can significantly influence the performance of the deep learning models employed for face recognition. Recognizing the importance of image quality, we explore the use of feature norms - the L2 norms of features extracted by a deep learning model - as a measure of image quality, addressing the gap in existing studies regarding the relationship between feature norms and image quality. We propose a novel hypothesis suggesting that samples with higher feature norms are more akin to the samples learned by the deep learning model, offering a new perspective on distinguishing noise samples in the dataset. Additionally, we introduce the Noise Direction Regularization (NDR) technique, a new regularization strategy that uses noise samples to improve face recognition performance in low-resolution settings. By integrating these insights and methodologies, our study aims to improve the accuracy and robustness of face recognition models, particularly in challenging environments where image quality varies widely.