Scholarlink

<p>Micro-expression recognition, a subfield of affective computing, plays a pivotal role in understanding genuine emotional states through subtle facial movements. Traditional methods have relied heavily on handcrafted features, which often fall short under complex conditions. Recent advancements in deep learning, particularly graph convolutional networks (GCN) and Transformers, have shown promise but face challenges in balancing local and global feature learning. To address this, we introduce a novel framework that integrates the multi-head spatial-aware self-attention (MHSSA) Transformer with GCN. This approach captures both spatial structures and temporal information in a 3D space, dynamically modeling features within and across predefined facial regions. The MHSSA-Transformer enhances fine-grained local feature extraction, while the GCN constructs adaptive adjacency matrices based on Action Unit co-occurrence and manifold distances, enabling more discriminative global feature learning. Experimental results on CASME II, SAMM, and SMIC datasets demonstrate state-of-the-art performance, with our method consistently outperforming existing approaches. This framework not only advances micro-expression recognition but also provides a robust solution for affective computing applications. The source code is available for research purposes at <a href="https://github.com/leyin20/MHSSA-Tranformer-GCN.git">https://github.com/leyin20/MHSSA-Tranformer-GCN.git</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

Integrating MHSSA-transformer and GCN for enhanced micro-expression recognition

  • 期刊论文

  • 发表日期：

    2026年2月20日

查看全文
  • Lin Zhu,
  • Chunlong Hu,
  • Huiru Zhao,
  • Hualong Yu

摘要

Micro-expression recognition, a subfield of affective computing, plays a pivotal role in understanding genuine emotional states through subtle facial movements. Traditional methods have relied heavily on handcrafted features, which often fall short under complex conditions. Recent advancements in deep learning, particularly graph convolutional networks (GCN) and Transformers, have shown promise but face challenges in balancing local and global feature learning. To address this, we introduce a novel framework that integrates the multi-head spatial-aware self-attention (MHSSA) Transformer with GCN. This approach captures both spatial structures and temporal information in a 3D space, dynamically modeling features within and across predefined facial regions. The MHSSA-Transformer enhances fine-grained local feature extraction, while the GCN constructs adaptive adjacency matrices based on Action Unit co-occurrence and manifold distances, enabling more discriminative global feature learning. Experimental results on CASME II, SAMM, and SMIC datasets demonstrate state-of-the-art performance, with our method consistently outperforming existing approaches. This framework not only advances micro-expression recognition but also provides a robust solution for affective computing applications. The source code is available for research purposes at https://github.com/leyin20/MHSSA-Tranformer-GCN.git.