Enhanced deep capsule network for EEG-based emotion recognition

Recently, it has become very popular to use electroencephalogram (EEG) signals in emotion recognition studies. But, EEG signals are much more complex than image and audio signals. There may be inconsistencies even in signals recorded from the same person. Therefore, EEG signals obtained from the human brain must be analyzed and processed accurately and consistently. In addition, traditional algorithms used to classify emotion ignore the neighborhood relationship and hierarchical order within the EEG signals. In this paper, a method including selection of suitable channels from EEG data, feature extraction by Welch power spectral density estimation of selected channels and enhanced capsule network-based classification model is presented. The most important innovation of the method is to adjust the architecture of the capsule network to adapt to the EEG signals. Thanks to the proposed method, 99.51% training and 98.21% test accuracy on positive, negative and neutral emotions were achieved in the Seed EEG dataset. The obtained results were also compared and evaluated with other state-of-the-art methods. Finally, the method was tested with Dreamer and Deap EEG datasets.