Scholarlink

<p>Deep learning has achieved remarkable success across multiple domains, but data labeling in practical applications is often affected by noisy labels, degrading model performance. On image datasets, such noise can mislead models into learning features irrelevant to the ground-truth labels, diverting them from the optimal convergence path. This paper proposes a training framework for noisy labels that combines label contrastive learning with dynamic label correction to enhance model robustness. Our framework first aligns image and label features via label contrastive learning, augmented by multi-view augmentation. Concurrently, it employs a dynamic noise selection strategy that identifies noisy labels based on dynamic scores and re-weights their loss contributions to mitigate adverse effects. Finally, it generates pseudo-labels through feature space alignment and iteratively updates sample weights. Experimental results show that this framework significantly improves performance on several benchmark datasets, especially under high noise ratio scenarios, by effectively integrating noisy sample selection with label correction strategies, reducing noise interference and improving model performance.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A noisy label learning method based on label contrastive learning and dynamic correction

  • 期刊论文

  • 发表日期：

    2026年4月8日

查看全文
  • Siyi Wang,
  • Junqi Xu,
  • Kunpeng Li,
  • Zengxu Wang

摘要

Deep learning has achieved remarkable success across multiple domains, but data labeling in practical applications is often affected by noisy labels, degrading model performance. On image datasets, such noise can mislead models into learning features irrelevant to the ground-truth labels, diverting them from the optimal convergence path. This paper proposes a training framework for noisy labels that combines label contrastive learning with dynamic label correction to enhance model robustness. Our framework first aligns image and label features via label contrastive learning, augmented by multi-view augmentation. Concurrently, it employs a dynamic noise selection strategy that identifies noisy labels based on dynamic scores and re-weights their loss contributions to mitigate adverse effects. Finally, it generates pseudo-labels through feature space alignment and iteratively updates sample weights. Experimental results show that this framework significantly improves performance on several benchmark datasets, especially under high noise ratio scenarios, by effectively integrating noisy sample selection with label correction strategies, reducing noise interference and improving model performance.