GLFNet: Attention Mechanism-Based Global-Local Feature Fusion Network for Micro-Expression Recognition

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Micro-expressions are subtle facial movements revealing true emotions.
• Micro-expression recognition (MER) is challenging due to short duration, low intensity, and imbalanced data.
• Existing MER methods struggle with effective feature extraction and data imbalance.

Purpose of the Study:

• To propose a novel Global-Local Feature Fusion Network (GLFNet) for enhanced MER.
• To address challenges of subtle movements, low intensity, and imbalanced datasets in MER.
• To improve the accuracy and robustness of micro-expression recognition systems.

Main Methods:

• Developed GLFNet with Global Attention (LA), Local Block (GB), and Adaptive Feature Fusion (AFF) modules.
• Employed attention mechanisms for global-local feature integration and salient local movement emphasis.
• Introduced a class-balanced loss function to mitigate dataset imbalance issues.

Main Results:

• GLFNet demonstrated superior performance over state-of-the-art methods on SMIC, CASME II, and SAMM datasets.
• Achieved significant improvements in unweighted F1-scores: 4.67% (SMIC), 2.02% (SAMM), 0.49% (CASME II).
• Validated the effectiveness of the global-local feature fusion strategy in MER.

Conclusions:

• The proposed GLFNet effectively extracts discriminative features for MER.
• The global-local feature fusion strategy significantly enhances micro-expression recognition accuracy.
• GLFNet represents a substantial advancement in the field of micro-expression analysis.