Emotion recognition based on feature weight analysis of multiple physiological signals.

Area of Science:

• Artificial Intelligence
• Cognitive Science
• Biomedical Engineering

Background:

• Emotion recognition from physiological signals is challenging due to unimodal limitations and difficulties in feature extraction.
• Existing deep learning methods often overlook channel importance and temporal dynamics in physiological data.

Purpose of the Study:

• To develop a multimodal deep learning framework for enhanced emotion recognition.
• To refine feature extraction from physiological signals using attention mechanisms.
• To improve the precision and generalization capability of emotion detection systems.

Main Methods:

• Extracted micro-differential-entropy (DE) emotion matrices from multi-channel EEG and peripheral physiological signals.
• Applied a channel-attention mechanism to reweight physiological signal importance across channels.
• Utilized depthwise-separable convolutional neural networks and long short-term memory networks for spatial and temporal feature extraction.
• Fused multimodal features and employed a multi-head attention mechanism within a recurrent network for temporal sequence analysis.

Main Results:

• The proposed multimodal framework demonstrated improved feature extraction from physiological signals.
• Attention mechanisms effectively captured channel importance and temporal dynamics.
• The approach achieved strong generalization capabilities on two distinct datasets.

Conclusions:

• The integrated multimodal framework with attention mechanisms significantly enhances emotion recognition accuracy.
• This method effectively addresses limitations of unimodal approaches by leveraging diverse physiological data.
• The findings suggest a promising direction for developing more robust and precise emotion detection technologies.