A Domain Generalization and Residual Network-Based Emotion Recognition from Physiological Signals

Area of Science:

• Physiological computing
• Affective computing
• Machine learning for signal processing

Background:

• Emotion recognition from physiological signals (ERPS) is challenging due to nonstationary, high-frequency data.
• Key challenges include capturing inter-signal correlations and addressing temporal covariate shift (TCS).

Purpose of the Study:

• To propose a domain generalization and residual network-based approach for ERPS (DGR-ERPS).
• To address limitations in feature extraction and temporal covariate shift in existing ERPS methods.

Main Methods:

• Pre-extracted time- and frequency-domain features were composed into new time series.
• Time series data were converted to 3D images for input into a residual-based feature encoder (RBFE).
• A domain generalization technique was employed to mitigate temporal covariate shift (TCS).

Main Results:

• The proposed DGR-ERPS method demonstrated superior performance in emotion recognition.
• Effectiveness was validated across both temporal covariate shift (TCS) and non-TCS scenarios.
• Extensive experiments on two real-world datasets confirmed the approach's robustness.

Conclusions:

• The DGR-ERPS approach effectively handles correlations and temporal covariate shift in physiological signals.
• This method offers a significant advancement for accurate and robust emotion recognition.
• The findings have broad implications for applications utilizing ERPS.