MCCA-VNet: A Vit-Based Deep Learning Approach for Micro-Expression Recognition Based on Facial Coding.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Psychology

Background:

• Micro-expressions offer more realistic insights than macro-expressions, valuable for psychological counseling and clinical diagnosis.
• Existing deep learning models, primarily using optical flow and Transformer, often overlook spatial relationships between facial landmarks.

Purpose of the Study:

• To propose MCCA-VNET, a deep learning model enhancing micro-expression recognition by incorporating spatial landmark relationships.
• To improve the accuracy and comprehensive performance of micro-expression identification.

Main Methods:

• Developed MCCA-VNET based on Transformer architecture.
• Integrated channel attention and spatial attention mechanisms into Vision Transformer.
• Extracted and fused changing facial features, emphasizing spatial and channel correlations.

Main Results:

• Achieved UF1 score of 0.8676 and UAR score of 0.8622 on a composite dataset (SAMM, CAS (ME) II, SMIC).
• Demonstrated superior performance across multiple indicators compared to previous state-of-the-art algorithms.
• Validated effectiveness through rigorous experimental testing on benchmark datasets.

Conclusions:

• MCCA-VNET significantly enhances micro-expression recognition accuracy.
• The model's fusion of attention mechanisms and spatial feature extraction offers a robust approach.
• The proposed method achieves the best comprehensive performance in micro-expression identification tasks.