Scholarlink

<p>Facial expression recognition (FER) encounters significant challenges under domain shifts caused by dataset bias and demographic variations. Existing cross-domain FER (CD-FER) methods typically rely on labeled source data, which may be restricted due to privacy concerns. To address this, we propose EERCL (Embedding Expression Relationship Contrastive Learning), a novel source-free CD-FER framework that adapts to target domains using only a pre-trained source model. By constructing an expression relationship matrix from source classifier weights, we design two novel loss functions: an expression relationship embedded class-aware contrastive loss (ER-CACo) and an expression relationship embedded instance discrimination contrastive loss (ER-IDCo). These losses enable effective adaptation without source data. Extensive experiments on multiple benchmark datasets demonstrate that EERCL outperforms existing source-free CD-FER methods and achieves competitive performance against state-of-the-art CD-FER approaches, with an average recognition accuracy of 66.84% across five target datasets. Our code and more details are available at <a href="https://github.com/Liusixone/EERCL_in_Source-Free.git.">https://github.com/Liusixone/EERCL_in_Source-Free.git.</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

Enhancing cross-domain facial expression recognition with expression relationship contrastive learning in a source-free setting

  • 期刊论文

  • 发表日期：

    2026年3月11日

查看全文
  • Zhe Guo,
  • Yi Liu,
  • Kui Wang,
  • Bingxin Wei,
  • Yi Wang

摘要

Facial expression recognition (FER) encounters significant challenges under domain shifts caused by dataset bias and demographic variations. Existing cross-domain FER (CD-FER) methods typically rely on labeled source data, which may be restricted due to privacy concerns. To address this, we propose EERCL (Embedding Expression Relationship Contrastive Learning), a novel source-free CD-FER framework that adapts to target domains using only a pre-trained source model. By constructing an expression relationship matrix from source classifier weights, we design two novel loss functions: an expression relationship embedded class-aware contrastive loss (ER-CACo) and an expression relationship embedded instance discrimination contrastive loss (ER-IDCo). These losses enable effective adaptation without source data. Extensive experiments on multiple benchmark datasets demonstrate that EERCL outperforms existing source-free CD-FER methods and achieves competitive performance against state-of-the-art CD-FER approaches, with an average recognition accuracy of 66.84% across five target datasets. Our code and more details are available at https://github.com/Liusixone/EERCL_in_Source-Free.git.