Automatic Emotion Recognition from EEG Signal Utilizing Wavelet Packet Node Reconstruction and a CNN Classifier

Emotion recognition using electroencephalogram (EEG) has encompassed a broad research area in the domain of affective computing since the change of emotional states in response to variations in brain activities. In this paper, a novel emotion recognition approach is proposed employing the wavelet decomposition and node reconstruction of EEG signals. Following the wavelet decomposition of the EEG signal into five different levels, a node reconstruction scheme is employed to reconstruct the signal at each level to retain the original signal length. Subsequently, a 2D matrix is formed by combining the reconstructed signal information at all decomposition levels. In view of incorporating the information of multi-channel EEG signal, a 3D frame is formed combining the 2D matrices acquired from all available channels. The frame is applied as an input to a 2D deep neural network where the signal information at different decomposition levels is passed to the depth dimension of the neural network. The proposed method offers substantial performance in the classification process for the valence and arousal domain considering the performance of all subjects. The average accuracies obtained for valence and arousal domains in binary class problems are 93.06% and 91.93% respectively.