Recognition of emotional states in response to audio-visual inductions based on nonlinear analysis and self-organisation map classification

Recently, emotion recognition using biological signals has attracted much attention by researchers due to the rapid development of machine learning algorithms and various applications of brain computer interface (BCI). This study addresses the emotion recognition system from electroencephalogram signals, in which different emotional states are represented on the valence and arousal dimensions. As regards to nonlinear nature and complex dynamics of EEG signals, we propose to use nonlinear features from brain electrical activity to evaluate emotional states. With this aim, we examined two different categories of nonlinear features: fractal-based features and entropy-based features. In addition to that, a two stage feature selection based on Dunn index and sequential forward feature selection (SFS) algorithm is employed for eliminating redundant and weak features, and finally SOM classifier was applied to selected features in order to classification of emotional classes. The experimental results show that the proposed method can represent user's emotional state effectively in both two-level and four-level of valence and arousal dimension. Furthermore, we determined the best channels and time segments for discriminating the emotions and the most related regions of brain to emotion-related sensory activities were found.