Scholarlink

<p>Micro-expression recognition (MER) is a challenging task due to the subtle and local movements of facial muscles. To get rid of redundant video frames, studies have been conducted on the use of apex frames for MER. However, these studies mainly focused on 2D frontal apex frames, ignoring side face information, which can result in insufficient extraction of features?. To address this issue, we propose a novel dual-stream network with coordinate attention (DSNCA) framework for MER, which comprises a multi-view coordinate attention module (MVCAM) and an apex frame coordinate attention module (AFCAM). The MVCAM initially employs 3D face reconstruction to acquire unobstructed multi-view images that are rich in micro-expression information and enhanced with geometric details. Subsequently, it adaptively learns micro-expression features from multiple angle views with coordinate attention, thus leveraging supplementary face information. In contrast, the AFCAM aims to adaptively locate key regions where micro-expressions occur, thereby minimizing redundant information. The proposed method achieved UF1 and UAR of 76.45, 75.16, 59.21, 59.6, 62.77, 62.06, 66.74, and 69.31 on the Chinese Academy of Sciences Micro-expression DatabaseII (CASMEII), the Spontaneous Micro-expression Corpus (SMIC), the Spontaneous Actions and Micro-Movements (SAMM), and the composite databases, respectively. The code is available for research purposes at <a href="https://github.com/pennypppp/DSNCA">https://github.com/pennypppp/DSNCA</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

Dual-stream network with coordinate attention for multi-view micro-expression recognition using 3D face reconstruction

  • 期刊论文

  • 发表日期：

    2025年9月8日

查看全文
  • Pianpian Ma,
  • Jingying Chen,
  • Xu Liu,
  • Xiaodi Liu

摘要

Micro-expression recognition (MER) is a challenging task due to the subtle and local movements of facial muscles. To get rid of redundant video frames, studies have been conducted on the use of apex frames for MER. However, these studies mainly focused on 2D frontal apex frames, ignoring side face information, which can result in insufficient extraction of features?. To address this issue, we propose a novel dual-stream network with coordinate attention (DSNCA) framework for MER, which comprises a multi-view coordinate attention module (MVCAM) and an apex frame coordinate attention module (AFCAM). The MVCAM initially employs 3D face reconstruction to acquire unobstructed multi-view images that are rich in micro-expression information and enhanced with geometric details. Subsequently, it adaptively learns micro-expression features from multiple angle views with coordinate attention, thus leveraging supplementary face information. In contrast, the AFCAM aims to adaptively locate key regions where micro-expressions occur, thereby minimizing redundant information. The proposed method achieved UF1 and UAR of 76.45, 75.16, 59.21, 59.6, 62.77, 62.06, 66.74, and 69.31 on the Chinese Academy of Sciences Micro-expression DatabaseII (CASMEII), the Spontaneous Micro-expression Corpus (SMIC), the Spontaneous Actions and Micro-Movements (SAMM), and the composite databases, respectively. The code is available for research purposes at https://github.com/pennypppp/DSNCA.