Emotion Recognition Based on EEG Signals and Deep Neural Networks Architectures

Emotion recognition is a significant field of study in the area of affective computing, with compelling applications for human-computer interaction and mental health. Various approaches have been proposed that are judged for their effectiveness, but there is the notion that physiological signals provide a more objective and reliable emotion estimation. In this study, we investigate the emotion recognition efficacy based on electroencephalogram (EEG) signals and machine learning techniques. We utilized four (4) publicly available datasets, the MAHNOB, SEED, SEED IV and DREAMER dataset to evaluate the genaralization performance of a Deep Neural Network (DNN) in classifying emotional states. Feature extraction techniques including time-domain, frequency-domain, and topographical maps (CSP) were utilized to derive feature representations of EEG signals. Using a simple Multi-Layer Perceptron (MLP) architecture we were able to achieve best accuracy of 94 % and best weighted F1 of 91 % for DREAMER arousal in Leave One Trial Out (LOTO) 93% and 90% for Leave One Subject Out (LOSO) respectively.