High Order Frequency Features as Emotion Discriminators

Considering that emotions are crucial for the correct interpretation of an individual’s actions, knowledge about the user’s emotional state is fundamental to make human-machine interaction (HMI) more natural. Countless works have sought to automate the process of recognizing emotions, being the analysis of brain signals obtained through electroencephalograms (EEG) one of the most used techniques. Knowing that EEG signals are non-stationary and contain relevant information in various frequency bands, analysis in the spectral domain is a good alternative to extract relevant features from these signals. In this context, HOS (Higher-order spectra) is an effective method for analyzing EEG. Therefore, the objective of this work was to develop a methodology for recognizing emotions through EEG signals using the HOS tool in the feature extraction stage. In this study, the signals from a database collected at the Federal University of São Francisco Valley were used to obtain images from HOS, referring to two classes of emotions. Afterwards, in the classification stage, those images were used to train and test a pre-trained convolutional neural network (CNN) VGG-19. The developed model reached an accuracy of 80.66%.