FERGCN: facial expression recognition based on graph convolution network

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Facial expression recognition is hindered by real-world complexities like occlusion, pose variation, and illumination changes.
• Existing methods struggle to robustly extract expression cues from unconstrained facial images.

Purpose of the Study:

• To develop a deep neural network, FERGCN, for effective facial expression recognition in challenging, in-the-wild environments.
• To improve the accuracy and robustness of facial expression analysis by leveraging both global and local facial features.

Main Methods:

• A novel deep neural network, FERGCN, integrating a feature extraction module (CNN with triplet attention and key point-guided attention) and a graph convolutional network.
• Utilizing a graph convolutional network to model correlations between global and local features based on facial key point topology.
• Incorporating a graph-matching module to enhance discrimination between different expressions using image similarity.

Main Results:

• FERGCN achieved high accuracy on public datasets: 88.23% on RAF-DB, 56.15% on SFEW, and 62.03% on AffectNet.
• The proposed method demonstrates effective facial expression recognition in real-world, complex scenarios.
• The integration of graph convolution and attention mechanisms significantly improved feature representation.

Conclusions:

• FERGCN offers a robust solution for facial expression recognition in unconstrained environments.
• The network's ability to handle occlusions and other challenges is significantly improved.
• This research contributes a powerful tool for advancing computer vision-based emotion analysis.