{Qiu Chen Filipe R. Cordeiro Yi Zhu Qian Zhang}

Abstract

Large-scale image datasets frequently contain unavoidable noisy labels, resulting in overfitting in deep neural networks and declining performance. Most existing methods for learning from noisy labels operate as one-stage frameworks, where training data division and semi-supervised learning (SSL) are intertwined for optimization. Accordingly, their effectiveness is significantly influenced by the precision of the separated clean set, prior knowledge of noise, and the robustness of SSL. In this paper, we propose a progressive sample selection framework with contrastive loss for noisy labels named PSSCL. This framework operates in two stages, using robust and contrastive losses to augment the robustness of the model. Stage I focuses on identifying a small clean set through a long-term confidence detection strategy, while stage II aims to enhance performance by expanding this clean set. PSSCL demonstrates significant improvement across various benchmarks when compared with state-of-the-art methods. The code is available at https://github.com/LanXiaoPang613/PSSCL.