MOTA, M. R. A.; http://lattes.cnpq.br/5372184836408953; MOTA, Moisés Roberto de Araújo.
Abstract:
Emotions are important for human beings, it influences our social behavior, memory and
cognition. The affective computing area aims to assist in the understanding of human emotional phenomena using computer systems for this. In this context, defining, charac-
terizing and classifying emotions and their constituent parts based on physiological signals, such as electroencephalogram signals (EEG), facial expressions, etc, is a challenging task. However, a optimal characteristics set for recognizing emotions from EEG signals is not yet known, although there are characteristics such as PSD (power spectral density) and HOC (higher-order crossings), widely used by the scientific community. New researches have substantially advanced the state of the art with regard to the recognition of emotions from EEG when considering the use of characteristics extracted from the phase space of these signals and classifiers based on deep learning techniques. In this context, the present thesis investigated the use of reconstruction of the EEG signals images of phase space, which are subjected to a 3D convolutional network (one of the deep learning techniques), to automatically learn the EEG signal space phase characteristics to recognize different emotional states independently of the individual. The results of the experiments demonstrated the feasibility and competitiveness of the proposed approach using images of the stacked phase space reconstruction of only one EEG FP1 channel, reaching accuracy of 0, 84 ± 0, 07, for four valence-arousal classes, and 0, 88 ± 0, 05 for two valence classes and 0, 94 ± 0, 01 for two arousal classes, both using the training and classification methodology of the LOSO type (leave-one-subject-out). These results, in addition to reducing the number of channels required for classification to the minimum possible, contribute to advancing the state of the art by presenting a new approach to the emotions classification based on EEG signals.