不确定性感知的标签噪声矫正算法

打开文本图片集

中图分类号：TP391 文献标志码：A DOI：10. 13705/j. issn.1671-6841.2024122

文章编号：1671-6841（2026）01-0010-09

Abstract： Label noise introduced issues of uncertainty into the training process of learning algorithms by reducing confidence in the prediction of true classes.To mitigate the impact of label noise，an uncertainty-aware label noise corrction （ULC）algorithm for robust classification was proposed.Firstly，based on evidence theory and subjective logic theory，uncertainty was estimated from multiple perspectives and the label information of the sample.Secondly，the dataset was finely divided into three subsets.The noise labels within these subsets were then corrected using joint prediction.Finally，to optimize the training objectives，each subset was processed using different regularization strategies. Comparative experiments were conducted on four simulated label noise datasets and two containing real label noise.On CIFAR-10 and CIFAR-100 with 40% pairflip-type label noise，the classification accuracy of ULC was increased by 10.58 percentage points and 15.84 percentage points compared to DivideMix，and the corrected label accuracy reached 95.48% and 81.32% ，respectively. The simulation results showed that the proposed algorithm accurately estimated uncertainty，finely improved the accuracy of corrected labels，and enhanced model generalization performance.

Key words： deep learning； label noise；uncertainty estimation； sample selection；label correction

0 引言

深度学习中的监督分类算法被应用于众多领域，大规模高质量数据集是其取得成功的关键。（剩余11818字）