Anti-noise face: A resilient model for face recognition with labeled noise data

Abstract

With the remarkable success of face recognition driven by large-scale datasets, noise learning has gained increasing attention due to the prevalence of noise within these datasets. While various margin-based loss functions and training strategies for label noise have been recently devised, two issues still remain to consider: (1) The explicit emphasis on specific characteristics of different types of noise is required. (2) The potential impact of noise during the early stages of training, which may lead to convergence issues, should not be ignored. In this study, we propose a comprehensive algorithm for learning with label noise. Compared to the existing noise self-correction methods, we further enhance detecting closed-set noise by introduce a closed-set noise self-correction module, and introduce a novel loss function for handling remaining noisy samples detected by an improved Gaussian Mixture Model. Additionally, we use a progressive approach, where we work through the easy examples first and then move on to the difficult ones, just as a student work through a course with the easy ones first and then the difficult ones later. Extensive experiments conducted on the synthesized noise dataset and on popular benchmarks have demonstrated the superior effectiveness of our approach over state-of-the-art alternatives.