Micro-Expression Recognition Using Uncertainty-Aware Magnification-Robust Networks.

Area of Science:

• Affective computing
• Pattern recognition
• Computer vision

Background:

• Micro-expressions (ME) are involuntary, subtle facial movements.
• Accurate micro-expression recognition (MER) has applications in interrogation and clinical diagnosis.
• Existing MER methods struggle with low ME intensity and imbalanced datasets.

Purpose of the Study:

• To propose a novel deep learning method, Uncertainty-Aware Magnification-Robust Networks (UAMRN), for enhanced MER.
• To address the challenges of subtle ME intensity and imbalanced sample distribution.
• To improve the accuracy and robustness of micro-expression recognition systems.

Main Methods:

• Implemented a sequence reconstruction technique to magnify subtle ME intensity.
• Introduced a Sparse Self-Attention (SSA) block, using Locality Sensitive Hashing (LSH) to suppress magnification artifacts.
• Guided network optimization using estimation confidence to address class imbalance, giving rare classes more attention.

Main Results:

• The UAMRN method demonstrated improved performance on three public ME databases (CASME II, SAMM, SMIC-HS).
• The proposed approach effectively handles low-intensity micro-expressions.
• The method shows superior results compared to recent state-of-the-art MER techniques.

Conclusions:

• UAMRN offers a straightforward and effective deep learning solution for MER.
• The method successfully tackles key challenges in micro-expression recognition.
• This work advances the field of affective computing and pattern recognition for facial expression analysis.