EEG-Based Emotion Recognition by Convolutional Neural Network with Multi-Scale Kernels.

Area of Science:

• Neuroscience
• Machine Learning
• Affective Computing

Background:

• Electroencephalogram (EEG) data offer a robust alternative for emotion recognition, overcoming limitations of deceptive external expressions.
• Traditional EEG emotion recognition methods often overlook crucial correlations between channels and frequency bands.
• Accurate feature extraction and delineation are vital for reliable EEG-based emotion prediction.

Purpose of the Study:

• To develop an advanced EEG analysis method for improved emotion recognition.
• To leverage the correlations between EEG channels and frequency bands for enhanced prediction accuracy.
• To utilize a 3D representation of EEG signals for learning local and global patterns.

Main Methods:

• Extracted time-domain features from 32-channel EEG signals and arranged them into feature-homogeneous matrices.
• Employed a 2D Convolutional Neural Network (CNN) with varying kernel sizes to capture local and global spatial-frequency patterns.
• Validated the approach using ten-fold cross-validation on the DEAP dataset.

Main Results:

• The proposed method achieved high average accuracies of 98.27% for arousal and 98.36% for valence binary classification.
• Incorporating channel and frequency band correlations significantly improved emotion prediction performance.
• The 2D CNN effectively learned patterns from the 3D EEG signal representation.

Conclusions:

• The novel EEG analysis method demonstrates superior performance in emotion recognition.
• Considering inter-channel and inter-frequency band correlations is crucial for accurate affective state prediction.
• The developed 2D CNN model provides an effective framework for analyzing complex EEG data patterns.