Scholarlink

<p>Facial expressions contain abundant emotional information, and micro-expressions (MEs), as a kind of spontaneous non-verbal emotional expression with short duration and subtle movement, objectively reflect the actual state that people try to hide. Therefore, MEs play an important role in many fields such as criminal trials and treatment of psychological diseases. Due to the low intensity and short duration of MEs, it is difficult to accurately distinguish MEs. To address this problem, we propose a ME recognition algorithm based on action unit (AU) and graph convolution network (GCN), building a spatio-temporal attention (STA) network for global and local AU information (GLAU) to recognize MEs, which can be called GLAU-STA. For ME segments, we first extract global facial features and encode the co-occurrence among local AUs. Then, the spatio-temporal attention network is built based on graph convolution and 1D convolution to process spatial features within frames and temporal features between frames, and self-attention is used to attach weights to different frames to improve the utilization of features and thereby improve the accuracy of the GLAU-STA. Experiments were performed on the SAMM dataset, and with the results showing that the GLAU-STA outperforms other algorithms in terms of effectiveness.</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

Global and local AU-assisted graph convolutional network for micro-expression recognition

  • 期刊论文

  • 发表日期：

    2025年5月27日

查看全文
  • Hongbo Xing,
  • Zhongjun Ma,
  • Bo Su,
  • Ting Cao,
  • Zhenxing Li,
  • Jie Wang,
  • Zilong Shao,
  • Yong Liu,
  • Yewen Cao

摘要

Facial expressions contain abundant emotional information, and micro-expressions (MEs), as a kind of spontaneous non-verbal emotional expression with short duration and subtle movement, objectively reflect the actual state that people try to hide. Therefore, MEs play an important role in many fields such as criminal trials and treatment of psychological diseases. Due to the low intensity and short duration of MEs, it is difficult to accurately distinguish MEs. To address this problem, we propose a ME recognition algorithm based on action unit (AU) and graph convolution network (GCN), building a spatio-temporal attention (STA) network for global and local AU information (GLAU) to recognize MEs, which can be called GLAU-STA. For ME segments, we first extract global facial features and encode the co-occurrence among local AUs. Then, the spatio-temporal attention network is built based on graph convolution and 1D convolution to process spatial features within frames and temporal features between frames, and self-attention is used to attach weights to different frames to improve the utilization of features and thereby improve the accuracy of the GLAU-STA. Experiments were performed on the SAMM dataset, and with the results showing that the GLAU-STA outperforms other algorithms in terms of effectiveness.