Emotion Recognition Method Using U-Net Neural Network With Multichannel EEG Features and Differential Entropy Characteristics

To improve accuracy in EEG-based emotion recognition and address current research limitations, providing a more robust framework for scientific study and mental health support. This paper presents an emotion recognition method tailored for small sample training data, effectively addressing the issue of emotion classification based on multichannel continuous electroencephalogram (EEG) signals. The proposed method first extracts differential entropy features from EEG signals and maps them into two-dimensional matrices based on their spectral characteristics. These preprocessed original data and five different frequency band matrices are then stacked into three-dimensional multichannel data, which are fed into the 3D-EEGU-Net model for classification. Experiments conducted on the public SEED dataset demonstrate an emotion recognition accuracy of 92.34%. Furthermore, to enhance the accuracy of emotion recognition for specific populations, the size of data blocks (patches) and model hyperparameters were dynamically adjusted. In the valence-based emotion ternary classification task conducted on the DEAP public dataset, the method achieved an emotion recognition accuracy of 72.02%. Comparative experiments show that the proposed method exhibits significant superiority in emotion recognition performance on both datasets. This research not only provides a new approach to EEG emotion recognition but also holds substantial potential for applications in detecting psychological disorders and physical conditions in individuals from specialized occupational groups.