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

Junnan Li,Jiang Li,Xiaoping Wang,Xin Zhan,Zhigang Zeng

DOI: https://doi.org/10.34133/cbsystems.0074

2024-01-01

Cyborg and Bionic Systems

Abstract:Emotion recognition from physiological signals (ERPS) has drawn tremendous attention and can be potentially applied to numerous fields. Since physiological signals are nonstationary time series with high sampling frequency, it is challenging to directly extract features from them. Additionally, there are 2 major challenges in ERPS: (a) how to adequately capture the correlations between physiological signals at different times and between different types of physiological signals and (b) how to effectively minimize the negative effect caused by temporal covariate shift (TCS). To tackle these problems, we propose a domain generalization and residual network-based approach for emotion recognition from physiological signals (DGR-ERPS). We first pre-extract time- and frequency-domain features from the original time series to compose a new time series. Then, in order to fully extract the correlation information of different physiological signals, these time series are converted into 3D image data to serve as input for a residual-based feature encoder (RBFE). In addition, we introduce a domain generalization-based technique to mitigate the issue posed by TCS. We have conducted extensive experiments on 2 real-world datasets, and the results indicate that our DGR-ERPS achieves superior performance under both TCS and non-TCS scenarios.