Emotion Recognition from Time-Frequency Analysis in EEG Signals Using a Deep Learning Strategy

Abstract

At recent years emerging technologies have experiences an exponential growing in the development of human computer interfaces with commercial or scientific purposes; in which is included the Affective computing. This area aims to develop computer interfaces to estimate human emotional states during an specific interaction environment with the goal of adapt itself to the users feelings in order to enhance their experience. In this paper a Deep learning based model is proposed to determine the emotional user state considering the Russell emotional model. This consist in determine high-low intensity (arousal) and positive-negative emotional intention (valence). The emotional recognition is based in time-frequency analysis through power spectral density and spectrogram features extracted from electroencephalography (EEG) signals in the well-known DEAP database. Besides, since Deep learning strategies require a significant amount of samples to effectively train their parameters; a volumetric framing consisting in signals windowing is employed. Finally, a non typical k-fold cross validation strategy is employed as most works in the state of the art; but a leave one subject group out (LOSGO) is used since it consist in a more suitable and reliable strategy for realistic interaction scenarios. Results for models to recognize emotional arousal and valence in EEG signals demonstrates efficiency to recognize emotions from different participants under different trials and affective stimuli.