A Transformer Convolutional Network With the Method of Image Segmentation for EEG-Based Emotion Recognition

Electroencephalogram (EEG) based emotion recognition has become an important topic in human-computer interaction and affective computing. However, existing advanced methods still have some problems. Firstly, using too many electrodes will decrease the practicality of EEG acquisition device. Secondly, transformer is not good at extracting local features. Finally, differential entropy (DE) is unsuitable for extracting features outside the 2–44 Hz frequency band. To solve these problems, we designed a neural network using 14 electrodes, utilizing differential entropy and designed spectrum sum (SS) to extract features, using convolutional neural networks and image segmentation techniques to learn local features, and transformer encoders to learn global features. The model outperformed advanced methods with classification results of 98.50% and 99.00% on the SEED-IV and SEED-V datasets.