A multi-scale deep CNN based on attention mechanism for EEG emotion recognition

Area of Science:

• Brain-Computer Interface (BCI)
• Affective Computing
• Neuroscience

Background:

• Emotion recognition is a key challenge in brain-computer interface (BCI) technology.
• Deep learning models show promise but require more discriminative features for EEG analysis.

Purpose of the Study:

• To develop an advanced deep learning model for enhanced EEG-based emotion recognition.
• To address challenges in extracting discriminative features from multi-channel EEG signals.

Main Methods:

• A multi-scale convolutional neural network (MSCNN) integrated with channel and spatial attention (CSA-MSCNN) was proposed.
• Channel attention refines feature extraction by weighting critical channels and reducing noise.
• Spatial attention precisely identifies emotion-related brain regions, while MSCNN captures local and deep signal features.

Main Results:

• CSA-MSCNN achieved high accuracies: 95.75% for valence and 95.39% for arousal (3-class) on the DEAP dataset.
• On the SEED dataset, it reached an average 3-class classification accuracy of 90.48% for emotion recognition.
• The model significantly outperformed traditional machine learning and showed strong competitiveness against graph convolutional neural networks (GCNN).

Conclusions:

• The CSA-MSCNN effectively handles the complexity of multi-channel EEG signals and regional information.
• This approach offers a robust solution for accurate and reliable EEG emotion recognition.