A Cascade Attention Based Facial Expression Recognition Network by Fusing Multi-Scale Spatio-Temporal Features.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Facial expression recognition (FER) systems struggle with performance degradation due to uneven illumination and partial face occlusions.
• Accurate identification of attention hotspots in dynamic facial regions (eyes, nose, mouth) is crucial for robust FER.
• Existing methods often fail to precisely capture spatial and temporal features under adverse conditions.

Purpose of the Study:

• To develop a novel cascade attention-based facial expression recognition network.
• To enhance the precision of attention to dynamic facial regions.
• To improve FER performance under challenging environmental conditions.

Main Methods:

• Proposed a hybrid approach combining an attention mechanism with pyramid features.
• Integrated local spatial features, multi-scale-stereoscopic spatial context features (from 3-scale pyramid), and temporal features.
• Developed a cascade attention-based network architecture.

Main Results:

• Achieved high recognition accuracy rates on benchmark datasets: CK+ (99.23%), Oulu-CASIA (89.29%), and RAF-DB (86.80%).
• Demonstrated superior performance compared to state-of-the-art methods.
• Validated effectiveness in both controlled experimental settings and natural environments.

Conclusions:

• The proposed cascade attention-based network effectively addresses limitations of current FER systems.
• The hybrid feature extraction and attention mechanism significantly improve recognition accuracy.
• The method shows strong generalization capabilities for real-world facial expression recognition.