Linking Multi-Layer Dynamical GCN With Style-Based Recalibration CNN for EEG-Based Emotion Recognition.

Area of Science:

• Neuroscience
• Computer Science
• Human-Computer Interaction

Background:

• Electroencephalography (EEG)-based emotion computing is a key area in human-computer interaction (HCI).
• Traditional convolutional neural networks (CNNs) struggle to capture complex brain network interactions crucial for emotion recognition.
• Understanding information transmission between neurons is vital for accurate emotion state analysis.

Purpose of the Study:

• To propose a novel model, MDGCN-SRCNN, that integrates graph convolutional networks (GCN) and CNNs for enhanced emotion recognition.
• To effectively learn channel connectivity and deep abstract features for distinguishing different emotional states.
• To improve the accuracy of emotion recognition by leveraging both shallow and deep layer features.

Main Methods:

• Developed a hybrid model (MDGCN-SRCNN) combining graph convolutional network and convolutional neural network architectures.
• Incorporated a style-based recalibration module within the CNN to extract emotion-relevant deep layer features.
• Conducted experiments on the SEED and SEED-IV datasets to validate the model's performance.

Main Results:

• Achieved high recognition accuracies of 95.08% on the SEED dataset and 85.52% on the SEED-IV dataset.
• Demonstrated superior performance compared to existing state-of-the-art methods.
• Visualization confirmed that combining shallow and deep layer features significantly enhances recognition performance.

Conclusions:

• The MDGCN-SRCNN model effectively captures brain network interactions for accurate emotion recognition.
• The proposed model outperforms current methods in EEG-based emotion computing.
• Analysis of connection weights identified important brain regions and channel relationships involved in emotion generation.