Distract Your Attention: Multi-Head Cross Attention Network for Facial Expression Recognition

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Facial expression recognition faces challenges due to subtle differences between classes and the need for holistic analysis of facial regions.
• Existing methods may struggle to capture high-order interactions among local facial features for accurate recognition.

Purpose of the Study:

• To propose a novel facial expression recognition network, the Distract your Attention Network (DAN), inspired by biological visual perception.
• To address the limitations of subtle appearance variations and the requirement for holistic feature interactions in facial expression recognition.

Main Methods:

• The Distract your Attention Network (DAN) integrates three components: Feature Clustering Network (FCN) for robust feature extraction, Multi-head Attention Network (MAN) for attending to multiple facial areas, and Attention Fusion Network (AFN) for fusing feature maps.
• FCN employs a large-margin learning objective to maximize class separability.
• MAN and AFN work synergistically to capture high-order interactions among local features across multiple facial regions.

Main Results:

• The proposed DAN method achieved state-of-the-art performance on three public facial expression recognition datasets: AffectNet, RAF-DB, and SFEW 2.0.
• Experiments demonstrated the effectiveness of DAN in recognizing subtle facial expressions by considering interactions across multiple facial regions.
• The DAN code is publicly available for further research and application.

Conclusions:

• The Distract your Attention Network (DAN) offers a significant advancement in facial expression recognition accuracy.
• The network's architecture, inspired by human visual perception, effectively handles subtle expression variations and holistic feature analysis.
• DAN represents a new state-of-the-art in automated facial expression recognition systems.