Spatio-temporal deep forest for emotion recognition based on facial electromyography signals

Area of Science:

• Human-Computer Interaction
• Biomedical Engineering
• Machine Learning

Background:

• Facial electromyogram (fEMG) is a crucial physiological signal for emotion recognition in human-computer interaction.
• Deep learning models show promise for fEMG-based emotion recognition but are limited by feature extraction challenges and large data requirements.

Purpose of the Study:

• To propose a novel spatio-temporal deep forest (STDF) model for classifying discrete emotions (neutral, sadness, fear) using multi-channel fEMG signals.
• To address limitations in feature extraction and training data scale for improved emotion recognition performance.

Main Methods:

• Developed a spatio-temporal deep forest (STDF) model integrating 2D frame sequences and multi-grained scanning for feature extraction.
• Employed a cascade forest-based classifier with adaptive layer structures for varying data scales.
• Evaluated the model on an in-house dataset of three-channel fEMG signals from twenty-seven subjects.

Main Results:

• The STDF model achieved a superior average accuracy of 97.41% in recognizing neutral, sadness, and fear emotions.
• The model demonstrated robustness by maintaining high accuracy even with a 50% reduction in training data scale.
• Significantly outperformed five other comparison methods in emotion recognition tasks.

Conclusions:

• The proposed STDF model offers an effective and efficient solution for fEMG-based emotion recognition.
• This approach enhances practical applicability by reducing reliance on large-scale datasets.
• Paves the way for more accessible and accurate emotion recognition systems in human-computer interaction.