An Improved Canonical Correlation Analysis for EEG Inter-Band Correlation Extraction

As the most active and promising research fields of affective computing, emotion recognition based on EEG signal has attracted much attention. Traditional methods usually pay their attention on single-channel features which reflect time-domain, or frequency-domain of EEG and bi-channel features which reflecting channel-wise relationship across brain regions. However, emotional features which capturing the coupling between the EEG frequency bands was seldom to discuss. In this paper, we proposed a method to extract the inter-bands correlation (IBC) features based on canonical correlation analysis (CCA). Firstly, we verified the validity of the IBC features through several experiments and found that the more correlated features between the EEG frequency bands contribute more to emotion classification. Then, the IBC features and traditional differential entropy (DE) were fused at the decision-level, which significantly improves the accuracy of emotion recognition on SEED dataset and local CUMULATE dataset. Our results show that IBC features is a promising method to promote the emotion recognition accuracy.