Scholarlink

<p>Micro expressions (MEs) are crucial for emotional communication. However, their short duration and subtle intensity make micro expression recognition (MER) a challenging task in computer vision. For MER, multi branch networks are often employed to address the challenge of limited sample sizes, but this approach also increases the network complexity and susceptibility to overfitting. To address these problems, we propose HCANet, which is an integrated convolution and attention network with half channel division for micro expression recognition. The network consists of a shallow detail enhancement convolution block (SDECBlock) and a half channel mixed convolution and attention block (HCABlock). Firstly, optical flow features are extracted from four regions of interest (the left eye, right eye, left lip, and right lip) and concatenated. Secondly, these features are processed through the SDECBlock, where reparameterization enables parallel differential convolutions. Finally, the resulting shallow features are passed through the HCABlock, which divides the features into two halves, multi scale features are extracted from one half via hybrid spatial channel attention (HSCA) and local features are extracted from the other half via deformable convolution (DConv). This dual processing reduces the overhead and enhances feature fusion. A large number of experimental results on MEGC2019 show that the <i>UF</i>1 and <i>UAR</i> metrics of the HCANet method reach 0.870 and 0.855, respectively, which are superior to those of the existing MER method. Thus, the proposed network can be widely applied in the fields of clinical diagnosis, public safety, sentiment analysis, and depression treatment.</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

HCANet: An integrated convolution and attention network with half channel division for micro expression recognition

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年11月10日

查看全文
  • Junying Gan,
  • Hantian Chen,
  • Zhenxin Zhuang,
  • Junling Xiong,
  • Wenchao Xu,
  • Zhen Chen

摘要

Micro expressions (MEs) are crucial for emotional communication. However, their short duration and subtle intensity make micro expression recognition (MER) a challenging task in computer vision. For MER, multi branch networks are often employed to address the challenge of limited sample sizes, but this approach also increases the network complexity and susceptibility to overfitting. To address these problems, we propose HCANet, which is an integrated convolution and attention network with half channel division for micro expression recognition. The network consists of a shallow detail enhancement convolution block (SDECBlock) and a half channel mixed convolution and attention block (HCABlock). Firstly, optical flow features are extracted from four regions of interest (the left eye, right eye, left lip, and right lip) and concatenated. Secondly, these features are processed through the SDECBlock, where reparameterization enables parallel differential convolutions. Finally, the resulting shallow features are passed through the HCABlock, which divides the features into two halves, multi scale features are extracted from one half via hybrid spatial channel attention (HSCA) and local features are extracted from the other half via deformable convolution (DConv). This dual processing reduces the overhead and enhances feature fusion. A large number of experimental results on MEGC2019 show that the UF1 and UAR metrics of the HCANet method reach 0.870 and 0.855, respectively, which are superior to those of the existing MER method. Thus, the proposed network can be widely applied in the fields of clinical diagnosis, public safety, sentiment analysis, and depression treatment.