Scholarlink

<p>Dynamic Facial Expression Recognition (DFER) necessitates the capture of subtle facial muscle movements and the modelling of their temporal evolution. Traditional methods often struggle with computational efficiency and long-range dependency modelling when processing long video sequences. This paper introduces CoESTNet, a cooperative spatio-temporal network, which employs a hierarchical design to address these challenges. CoESTNet integrates a ResNet-18 for initial spatial feature extraction, a Multi-Scale Feature Enhancement module for detailed spatial refinement, a Temporal Feature Aggregation Gated (TFAG) module for temporal saliency distillation via adaptive frame-importance weighting, and a lightweight Transformer for global context modelling. Notably, the TFAG-learned importance weights are reused to encourage cooperation between temporal distillation and global dependency modelling. Experiments on DFEW, CK + , and Oulu-CASIA benchmarks demonstrate CoESTNet’s superior accuracy-efficiency trade-off, achieving a weighted average recall (WAR) of 69.58% on DFEW and outperforming state-of-the-art methods. Here, we show that our approach significantly enhances DFER performance, providing a practical solution for real-world applications.</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 dynamic facial expression recognition through cooperative spatio-temporal feature learning

  • 期刊论文

  • 发表日期：

    2026年4月28日

查看全文
  • Chunyan Yu,
  • Meiling Liu

摘要

Dynamic Facial Expression Recognition (DFER) necessitates the capture of subtle facial muscle movements and the modelling of their temporal evolution. Traditional methods often struggle with computational efficiency and long-range dependency modelling when processing long video sequences. This paper introduces CoESTNet, a cooperative spatio-temporal network, which employs a hierarchical design to address these challenges. CoESTNet integrates a ResNet-18 for initial spatial feature extraction, a Multi-Scale Feature Enhancement module for detailed spatial refinement, a Temporal Feature Aggregation Gated (TFAG) module for temporal saliency distillation via adaptive frame-importance weighting, and a lightweight Transformer for global context modelling. Notably, the TFAG-learned importance weights are reused to encourage cooperation between temporal distillation and global dependency modelling. Experiments on DFEW, CK + , and Oulu-CASIA benchmarks demonstrate CoESTNet’s superior accuracy-efficiency trade-off, achieving a weighted average recall (WAR) of 69.58% on DFEW and outperforming state-of-the-art methods. Here, we show that our approach significantly enhances DFER performance, providing a practical solution for real-world applications.