Analysis of Spatio-temporal Relationship of Multiple Energy Spectra of EEG Data for Emotion Recognition

Evaluation of several feature metrics derived from decomposed wavelet coefficients of electroencephalographic data for emotion recognition is presented in this paper. Five different emotions (joy, sadness, disgust, fear, and neutral) are elicited by providing stimulus patterns and EEG data is recorded for each participant. The collected dataset is preprocessed through a band-pass filter, a notch filter, and a Laplacian Montage for noise and artifact removal. Discrete Wavelet Transform based spectral decomposition is employed to separate each of the 256-channel data into 5 specific frequency bands (Delta, Theta, Alpha, Beta, and Gamma bands), and several feature metrics are calculated to represent different emotions. A multi-layer perceptron neural network is used to classify the feature data into different emotions. Experimental evaluations performed on EEG data captured by a 256 channel EGI data acquisition system shows promising results with an average emotion recognition rate of 91.73% for 5 subjects.