Hybrid Attention Cascade Network for Facial Expression Recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Facial expression recognition is crucial for human-computer interaction.
• Existing methods struggle with real-world variations like illumination and head pose.
• The Acted Facial Expressions in the wild (AFEW) dataset presents significant challenges.

Purpose of the Study:

• To propose a novel cascade network for improved facial expression recognition.
• To enhance performance on the challenging AFEW dataset.
• To achieve competitive accuracy comparable to state-of-the-art methods.

Main Methods:

• A cascade network integrating spatial feature extraction, hybrid attention, and temporal feature extraction.
• Face detection, ROI extraction, and facial landmark-based alignment.
• Utilizing a residual neural network for spatial features and a gate control loop unit for temporal features.

Main Results:

• Achieved 98.46% accuracy on CK+, 87.31% on Oulu-CASIA, and 53.44% on AFEW.
• Demonstrated a performance improvement exceeding 2% on the AFEW dataset.
• Showcased competitive performance against state-of-the-art facial expression recognition methods.

Conclusions:

• The proposed cascade network effectively extracts spatial and temporal features for robust facial expression recognition.
• The method shows significant outperformance in natural environments, addressing real-world constraints.
• This approach offers a promising direction for advancing emotion recognition technology.