Scholarlink

<p>Micro-expression recognition (MER) analyzes tiny, involuntary facial muscle movements to identify an individual’s emotions. Recently, several deep learning-based approaches have confirmed that spatiotemporal features can effectively improve the performance of MER models. However, these methods cannot effectively utilize the temporal information of micro-expressions (MEs) existing in specific facial regions through prior knowledge. To address this issue, this paper proposes a <b>M</b>icro-<b>E</b>xpression <b>R</b>ecognition via <b>D</b>ual-<b>S</b>tream <b>L</b>RCN-<b>E</b>nhanced <b>S</b>patiotemporal <b>A</b>ttention <b>N</b>etwork (LRCN-STAN) that integrates Convolutional Block Attention Module (CBAM) and Long Short-Term Memory (LSTM) for feature fusion to capture the detailed spatiotemporal features of MEs. The LRCN-STAN consists of two components: Optical Flow-based Attentional Feature Extraction (OFAFE) and Longitudinal Video Sequence Feature Extraction (LVSFE). For OFAFE, a Lightweight Residual Convolutional block network (LRCN) based on the CBAM is maintained to enhance the learning of crucial spatial information in the flow feature. For LVSFE, a combined LRCN and LSTM is employed for temporal feature extraction. Attention regions are handcrafted to enhance subtle motions in ME videos using prior knowledge to effectively filter out background noise and irrelevant information. Ultimately, the Information Noise Contrastive Estimation (InfoNCE) loss is leveraged for intermediate supervision to ensure that the different features extracted by the two streams are effectively fused. Extensive experiments on the CASME II, SMIC, SAMM, and Composite datasets demonstrate that LRCN-STAN can effectively cope with the recognition of MEs.</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

LRCN-STAN: micro-expression recognition via dual-stream LRCN-enhanced spatiotemporal attention network

  • 期刊论文

  • 发表日期：

    2026年1月29日

查看全文
  • Yujie Zhang,
  • Xiaofang Guo,
  • Siyu Wu,
  • Wenjie Xu,
  • Yifan Zheng

摘要

Micro-expression recognition (MER) analyzes tiny, involuntary facial muscle movements to identify an individual’s emotions. Recently, several deep learning-based approaches have confirmed that spatiotemporal features can effectively improve the performance of MER models. However, these methods cannot effectively utilize the temporal information of micro-expressions (MEs) existing in specific facial regions through prior knowledge. To address this issue, this paper proposes a Micro-Expression Recognition via Dual-Stream LRCN-Enhanced Spatiotemporal Attention Network (LRCN-STAN) that integrates Convolutional Block Attention Module (CBAM) and Long Short-Term Memory (LSTM) for feature fusion to capture the detailed spatiotemporal features of MEs. The LRCN-STAN consists of two components: Optical Flow-based Attentional Feature Extraction (OFAFE) and Longitudinal Video Sequence Feature Extraction (LVSFE). For OFAFE, a Lightweight Residual Convolutional block network (LRCN) based on the CBAM is maintained to enhance the learning of crucial spatial information in the flow feature. For LVSFE, a combined LRCN and LSTM is employed for temporal feature extraction. Attention regions are handcrafted to enhance subtle motions in ME videos using prior knowledge to effectively filter out background noise and irrelevant information. Ultimately, the Information Noise Contrastive Estimation (InfoNCE) loss is leveraged for intermediate supervision to ensure that the different features extracted by the two streams are effectively fused. Extensive experiments on the CASME II, SMIC, SAMM, and Composite datasets demonstrate that LRCN-STAN can effectively cope with the recognition of MEs.