CATM: A Multi-Feature-Based Cross-Scale Attentional Convolutional EEG Emotion Recognition Model.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Signal Processing

Background:

• Existing electroencephalogram (EEG) emotion recognition methods often fail to fully exploit temporal, frequency, and spatial information, leading to suboptimal classification accuracy.
• This limitation hinders the development of effective brain-computer interfaces and affective computing applications.

Purpose of the Study:

• To propose a novel Convolutional Attention model (CATM) for EEG-based emotion recognition that integrates multi-feature and multi-frequency band information.
• To enhance the accuracy of EEG emotion classification by leveraging cross-scale attention mechanisms.

Main Methods:

• Developed a Convolutional Attention model (CATM) incorporating a cross-scale attention module, frequency-space attention module, feature transition module, temporal feature extraction module, and a depth classification module.
• Extracted spatial features at different scales and assigned weights to important channels and spatial locations using attention mechanisms.
• Extracted temporal features and performed depth classification on preprocessed EEG signals.

Main Results:

• Achieved high accuracies on the DEAP dataset: 99.70% for valence and 99.74% for arousal in binary classification, and 97.27% in four-class classification.
• Demonstrated strong performance in few-channel (5-channel) experiments, with accuracies of 97.96% (valence) and 98.11% (arousal) for binary classification, and 92.86% for four-class classification.
• Outperformed other recent methods in both full-channel and few-channel EEG emotion recognition tasks.

Conclusions:

• The proposed CATM model effectively utilizes multi-domain information (time, frequency, space) for superior EEG emotion recognition.
• The model demonstrates robustness and high accuracy even with a reduced number of EEG channels, indicating its practical applicability.