Scholarlink

<p>Facial expression recognition (FER) is pivotal in interpreting human emotions and intentions, with applications spanning human–computer interaction, online monitoring, and healthcare. However, the presence of noisy labels in FER datasets, stemming from ambiguous facial expressions, poor image quality, and subjective annotator biases, poses significant challenges. Additionally, the high interclass similarity of facial expressions exacerbates these difficulties. To address these issues, we introduced the noisy label facial expression recognition (NLFER) network, which utilizes dynamic label updates to suppress noisy labels and prevent overfitting to uncertain images, and employs a multi-branch feature cross-fusion strategy to solve the problem of high similarity among classes. The NLFER architecture comprises a multi-branch feature cross-fusion module and a label ranking update module. The former integrates facial and landmark features to focus on salient facial regions, while the latter reweights samples and applies a threshold to regularize weights, reassigning labels to samples with predicted probabilities exceeding the threshold boundary. The experimental results show that NLFER outperforms other methods on the RAF-DB, FERPlus, and AffectNet datasets. Its accuracy rates are 89.21%, 89.06%, and 63.63%, respectively, all of which are superior to the existing methods. Our code is available at <a href="https://github.com/CYChe1/NLFER">https://github.com/CYChe1/NLFER</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

NLFER: multi-branch attention cross-fusion for robust facial expression recognition amidst noisy labels

  • 期刊论文

  • 发表日期：

    2025年12月1日

查看全文
  • Cheng-Yue Che,
  • Hong-Mei Sun,
  • Yu-Xiang Chen,
  • Shuang Yang,
  • Rui-Sheng Jia

摘要

Facial expression recognition (FER) is pivotal in interpreting human emotions and intentions, with applications spanning human–computer interaction, online monitoring, and healthcare. However, the presence of noisy labels in FER datasets, stemming from ambiguous facial expressions, poor image quality, and subjective annotator biases, poses significant challenges. Additionally, the high interclass similarity of facial expressions exacerbates these difficulties. To address these issues, we introduced the noisy label facial expression recognition (NLFER) network, which utilizes dynamic label updates to suppress noisy labels and prevent overfitting to uncertain images, and employs a multi-branch feature cross-fusion strategy to solve the problem of high similarity among classes. The NLFER architecture comprises a multi-branch feature cross-fusion module and a label ranking update module. The former integrates facial and landmark features to focus on salient facial regions, while the latter reweights samples and applies a threshold to regularize weights, reassigning labels to samples with predicted probabilities exceeding the threshold boundary. The experimental results show that NLFER outperforms other methods on the RAF-DB, FERPlus, and AffectNet datasets. Its accuracy rates are 89.21%, 89.06%, and 63.63%, respectively, all of which are superior to the existing methods. Our code is available at https://github.com/CYChe1/NLFER.