Scholarlink

<p>In this paper, we introduce a groundbreaking Dynamic region feature network (DRFN) tailored for end-to-end facial expression recognition (FER). Our approach views facial expressions as intricate combinations of diverse facial muscle movements across various regions. To accurately capture these subtle muscular shifts, DRFN dynamically learns multiple region-specific features from a shared holistic feature map. Furthermore, it integrates an advanced attention mechanism to assess the importance weight of each region, thereby refining the distinctive features while competing for a common resource pool. This innovative design allows DRFN to extract diverse and discriminant information from a broader range of facial regions, particularly when recognizing facial expressions becomes challenging due to minimal facial muscle movements. The visualization results demonstrate our method’s effectiveness in capturing nuanced facial cues. Consequently, leveraging the high-quality features extracted by DRFN, our approach achieves remarkable accuracy rates of 91.04% on the FERPlus dataset, 87.86% on the RAF-DB dataset, and 59.93% on the large-scale and highly complex AffectNet dataset. These results underscore the significant potential and practical value of our proposed model for real-world FER applications, setting a new benchmark in the field of artificial intelligence-driven facial expression recognition.</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

Dynamic region features learning for facial expression recognition

  • 期刊论文

  • 发表日期：

    2025年4月16日

查看全文
  • Yuanlun Xie,
  • Wenhao Wang,
  • Yibo Zhang,
  • Kaibo Shi,
  • Hengxin Zhang,
  • Nan Zhou

摘要

In this paper, we introduce a groundbreaking Dynamic region feature network (DRFN) tailored for end-to-end facial expression recognition (FER). Our approach views facial expressions as intricate combinations of diverse facial muscle movements across various regions. To accurately capture these subtle muscular shifts, DRFN dynamically learns multiple region-specific features from a shared holistic feature map. Furthermore, it integrates an advanced attention mechanism to assess the importance weight of each region, thereby refining the distinctive features while competing for a common resource pool. This innovative design allows DRFN to extract diverse and discriminant information from a broader range of facial regions, particularly when recognizing facial expressions becomes challenging due to minimal facial muscle movements. The visualization results demonstrate our method’s effectiveness in capturing nuanced facial cues. Consequently, leveraging the high-quality features extracted by DRFN, our approach achieves remarkable accuracy rates of 91.04% on the FERPlus dataset, 87.86% on the RAF-DB dataset, and 59.93% on the large-scale and highly complex AffectNet dataset. These results underscore the significant potential and practical value of our proposed model for real-world FER applications, setting a new benchmark in the field of artificial intelligence-driven facial expression recognition.