Temporal Capsule Feature Network for Eye-Tracking Emotion Recognition.

Area of Science:

• Physiological signal analysis
• Affective computing
• Human-computer interaction

Background:

• Eye tracking (ET) parameters are valuable physiological signals for emotion recognition.
• Current methods struggle with extracting temporal dynamics and emotionally specific features from ET data.
• Challenges include limited model robustness and individual generalization capabilities.

Purpose of the Study:

• To propose a novel Temporal Capsule Feature Network (TCFN) for enhanced eye tracking-based emotion recognition.
• To address limitations in temporal dynamic information extraction and feature specificity in existing models.
• To improve the robustness and individual generalization of emotion recognition systems.

Main Methods:

• Developed a Temporal Capsule Feature Network (TCFN) incorporating a Window Feature Module for temporal dynamics.
• Utilized a specialized Capsule Network Module to capture feature interdependencies.
• Implemented an MLP Classification Module and a Dual-Loss Mechanism for optimized performance.

Main Results:

• Achieved 83.27% average accuracy for Arousal and 89.94% for Valence (three-class) on the eSEE-d dataset.
• Reached 63.85% accuracy for four-category, across-session emotion recognition on the SEED-IV dataset.
• Demonstrated the superiority of the TCFN model in emotion recognition tasks.

Conclusions:

• The proposed TCFN effectively extracts temporal dynamic information and emotionally specific features from eye tracking data.
• TCFN shows significant improvements in accuracy and generalization for emotion recognition.
• This approach offers a promising direction for advanced affective computing systems.