Scholarlink

<p>Facial expression recognition, editing, and synthesis are crucial tasks with wide applications. However, existing methods often struggle with the visual readability and intelligibility of emotional states. To address this, we introduce a codec dual-output adversarial network (CDAN) for parsing facial expression patterns. Leveraging self-supervised learning and facial action units (AUs), CDAN distinguishes between expression and non-expression regions, generating global facial expression patterns. The spatial extend attention (SEA) module captures facial muscular fluctuations, while muscular textural and magnitude losses ensure convergence to realistic facial expressions. Experiments on AffectNet, KDEF, and Oulu-CASIA datasets demonstrate the effectiveness of CDAN, with a parsed ratio (PR) of 0.845. Our work not only defines facial expression parsing but also presents a novel framework that significantly enhances the readability and intelligibility of facial expressions. Code is available at <a href="https://github.com/ZackyCaleb/CDAN">https://github.com/ZackyCaleb/CDAN</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

Self-supervised facial expression parsing: unveiling global patterns through facial action units

  • 期刊论文

  • 发表日期：

    2025年4月14日

查看全文
  • Kai Zou,
  • Lianghai Jin,
  • Xiangyang Xu

摘要

Facial expression recognition, editing, and synthesis are crucial tasks with wide applications. However, existing methods often struggle with the visual readability and intelligibility of emotional states. To address this, we introduce a codec dual-output adversarial network (CDAN) for parsing facial expression patterns. Leveraging self-supervised learning and facial action units (AUs), CDAN distinguishes between expression and non-expression regions, generating global facial expression patterns. The spatial extend attention (SEA) module captures facial muscular fluctuations, while muscular textural and magnitude losses ensure convergence to realistic facial expressions. Experiments on AffectNet, KDEF, and Oulu-CASIA datasets demonstrate the effectiveness of CDAN, with a parsed ratio (PR) of 0.845. Our work not only defines facial expression parsing but also presents a novel framework that significantly enhances the readability and intelligibility of facial expressions. Code is available at https://github.com/ZackyCaleb/CDAN.