Dimensional Emotion Recognition Using EEG Signals via 1D Convolutional Neural Network

Emotion recognition using Electroencephalography (EEG) signals is an interesting and booming area of research in the field of human-computer interaction systems. EEG signals predict the emotions precisely because they experience the immediate effects of spontaneous emotion change without any external intervention. In this paper, we propose a deep learning algorithm, 1D convolutional neural network for determining dimensionality of emotions by effectively learning temporal features of time series-based EEG signals. The proposed model removes the artifacts using bandpass filtering and converts the signals into n number of 2D array segments. For all 32-channel data of EEG signals, 1D CNN network architecture with four 1D convolutional layers performs feature learning and classification. Experiments on MAHNOB-HCI dataset have demonstrated that the model achieves high accuracy 86.97% and 87.07% for valence and arousal, respectively, and outperform LSTM-RNN and GRU-RNN. Also, proposed 1D CNN uses only 13,664 trainable parameters and 675 ms running time to predict the dimensionality of emotion. It is observed that the proposed model gives significantly better efficiency for computational and operational cost as compared to LSTM-RNN and GRU-RNN network architectures.