Scholarlink

<p>Source-free unsupervised domain adaptation (SFUDA) is a pivotal challenge in machine learning, aiming to adapt models from labeled source domains to unlabeled target domains without access to source data. Existing pseudo-labeling methods often suffer from label noise and limited model generalization due to static augmentation strategies, and some methods discard low-confidence samples, leading to incomplete utilization of the target domain. To address these issues, we introduce a novel method named feature mixing-driven dynamic contrastive consistency learning (MDCC). MDCC employs a divide-and-conquer strategy, handling confident and non-confident subsets differently while preserving all target samples. At the core of MDCC, a feature mixing (FM) module performs feature-level augmentation by mixing target features with dynamically updated class centroids. For confident subsets, we propose feature mixing-driven contrastive learning (FMCL) to learn robust domain-invariant representations. For non-confident subsets, we introduce Dynamic Consistency Learning (DCL), which imposes regularization constraints without directly enforcing pseudo-label supervision. Extensive experiments on three domain adaptation benchmarks demonstrate that our method achieves competitive or superior recognition performance compared to state-of-the-art SFUDA approaches. Code: <a href="https://github.com/ABlackx/MDCC">https://github.com/ABlackx/MDCC</a>; <a href="https://doi.org/10.5281/zenodo.16892195">https://doi.org/10.5281/zenodo.16892195</a>.</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

Feature mixing-driven dynamic contrastive learning for robust source-free unsupervised domain adaptation

  • 期刊论文

  • 发表日期：

    2026年6月22日

查看全文
  • Junjie Chen,
  • Farong Gao,
  • Zhangyi Yang,
  • Qizhong Zhang

摘要

Source-free unsupervised domain adaptation (SFUDA) is a pivotal challenge in machine learning, aiming to adapt models from labeled source domains to unlabeled target domains without access to source data. Existing pseudo-labeling methods often suffer from label noise and limited model generalization due to static augmentation strategies, and some methods discard low-confidence samples, leading to incomplete utilization of the target domain. To address these issues, we introduce a novel method named feature mixing-driven dynamic contrastive consistency learning (MDCC). MDCC employs a divide-and-conquer strategy, handling confident and non-confident subsets differently while preserving all target samples. At the core of MDCC, a feature mixing (FM) module performs feature-level augmentation by mixing target features with dynamically updated class centroids. For confident subsets, we propose feature mixing-driven contrastive learning (FMCL) to learn robust domain-invariant representations. For non-confident subsets, we introduce Dynamic Consistency Learning (DCL), which imposes regularization constraints without directly enforcing pseudo-label supervision. Extensive experiments on three domain adaptation benchmarks demonstrate that our method achieves competitive or superior recognition performance compared to state-of-the-art SFUDA approaches. Code: https://github.com/ABlackx/MDCC; https://doi.org/10.5281/zenodo.16892195.