Learning Dynamic Relationships for Facial Expression Recognition Based on Graph Convolutional Network.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Facial Action Units (AUs) analysis is crucial for Facial Expression Recognition (FER).
• Existing Deep Spectral Convolutional Networks (DSCNs) use predefined regions and graphs, limiting dynamic spatial dependency modeling for FER.
• Current methods struggle to capture the close relationships between facial regions and specific AUs.

Purpose of the Study:

• To propose a novel Double Dynamic Relationships Graph Convolutional Network (DDRGCN) for enhanced FER.
• To dynamically model spatial dependencies between facial regions relevant to AUs.
• To develop a lightweight and efficient FER model.

Main Methods:

• Constructed facial graph data using 20 Regions of Interest (ROIs) guided by facial AUs.
• Developed a trainable weighted adjacency matrix to learn edge strengths in the facial graph.
• Implemented an efficient graph convolutional network for automatic learning of vertex dependencies.

Main Results:

• The DDRGCN model achieved superior accuracy on four widely-used FER datasets.
• Demonstrated significant reductions in model parameters (110K) and size (0.48MB) compared to existing methods.
• Showcased improved speed in facial expression estimation.

Conclusions:

• The proposed DDRGCN effectively models dynamic spatial dependencies for FER.
• Achieved state-of-the-art performance with a significantly smaller and faster model.
• Offers a promising lightweight solution for real-world facial expression recognition applications.