Combining Fixed-Weight ArcFace Loss and Vision Transformer for Facial Expression Recognition

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Facial expression recognition (FER) is challenged by intra-class variations.
• Existing methods like ArcFace loss enhance separability but don't constrain class center distribution.
• Deep learning, particularly Vision Transformers (ViTs), shows promise in feature learning for FER.

Purpose of the Study:

• To introduce a novel weight-constrained ArcFace loss function.
• To integrate this loss function into the Vision Transformer (ViT) framework for improved FER.
• To address biases from imbalanced data and reduce computational overhead.

Main Methods:

• Developed a weight-constrained ArcFace loss function.
• Integrated the proposed loss into a Vision Transformer (ViT) architecture.
• Evaluated performance on RAF-DB and FER2013 datasets, comparing against standard ArcFace and other methods.

Main Results:

• The proposed weight-constrained ArcFace loss significantly improved facial expression recognition accuracy.
• The approach demonstrated enhanced computational efficiency compared to standard ArcFace loss.
• Stabilized weight optimization led to reduced computational overhead and alleviated data distribution biases.

Conclusions:

• The weight-constrained ArcFace loss offers a more effective and efficient approach for facial expression recognition using ViTs.
• This method enhances model robustness against data imbalances and improves overall performance.
• The findings suggest a promising direction for advancing deep learning-based FER systems.