Scholarlink

<p>Micro-expression can reveal a person’s true feelings and possesses significant importance in fields such as criminal interrogation and psychological counseling. However, due to the subtlety and complexity of micro-expression, comprehensively understanding its features remains a considerable challenge. To address this challenge, this paper proposes a dual-branch network structure that integrates traditional optical flow with deep learning-based optical flow. The micro-expression features are extracted and processed in separate branches, thereby harnessing the complementary advantages of the two optical flow methods. The first branch employs the proposed Multi-Scale Patch Attention Convolution Network (MPACNet), which is designed to process Farneback optical flow by capturing local details. The second branch utilizes the Swin Transformer network with FlowNet2 optical flow, demonstrating outstanding performance in extracting global dynamic information. In addition, this framework effectively combines local information from traditional convolutional networks with global information from the Swin Transformer, achieving multi-stage feature fusion. Following the standards of Comprehensive Database Evaluation (CDE) and Single Database Evaluation (SDE), extensive experiments have been conducted on four datasets—SMIC-HS, CASME II, SAMM, and CAS(ME)<sup>3</sup>. The experimental results show that the proposed method demonstrates strong competitiveness across various evaluation metrics.</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

A dual-branch approach with multi-stage semantic integration and dual optical flow for micro-expression recognition

  • 期刊论文

  • 发表日期：

    2025年8月21日

查看全文
  • Shuhuan Zhao,
  • Peijing Zhao,
  • Zixin Hao,
  • Shuaiqi Liu

摘要

Micro-expression can reveal a person’s true feelings and possesses significant importance in fields such as criminal interrogation and psychological counseling. However, due to the subtlety and complexity of micro-expression, comprehensively understanding its features remains a considerable challenge. To address this challenge, this paper proposes a dual-branch network structure that integrates traditional optical flow with deep learning-based optical flow. The micro-expression features are extracted and processed in separate branches, thereby harnessing the complementary advantages of the two optical flow methods. The first branch employs the proposed Multi-Scale Patch Attention Convolution Network (MPACNet), which is designed to process Farneback optical flow by capturing local details. The second branch utilizes the Swin Transformer network with FlowNet2 optical flow, demonstrating outstanding performance in extracting global dynamic information. In addition, this framework effectively combines local information from traditional convolutional networks with global information from the Swin Transformer, achieving multi-stage feature fusion. Following the standards of Comprehensive Database Evaluation (CDE) and Single Database Evaluation (SDE), extensive experiments have been conducted on four datasets—SMIC-HS, CASME II, SAMM, and CAS(ME)3. The experimental results show that the proposed method demonstrates strong competitiveness across various evaluation metrics.