Parallel Sequence-Channel Projection Convolutional Neural Network for EEG-Based Emotion Recognition

One of the challenges in emotion recognition is finding an effective way to represent spatial-temporal features from EEG. To fully utilize the features on multiple dimensions of EEG signals, we propose a parallel sequence-channel projection convolutional neural network, including temporal stream sub-network, spatial stream sub-network, and fusion classification block. Temporal stream extracts temporal continuity via sequence-projection layer while spatial stream captures spatial correlation via channel-projection layer. Both sequence-projection and channel-projection adopt length-synchronized convolutional kernel to decode whole time and space information. The size of length-synchronized convolutional kernel is equal to the length of transmitted EEG sequence. The fusion classification block combines the extracted temporal and spatial features into a joint spatial-temporal feature vector for emotion prediction. In addition, we present a baseline noise filtering module to amplify input signals and a random channels exchange strategy to enrich the baseline-removed emotional signals. Experimental evaluation on DEAP dataset reveals that the proposed method achieves state-of-the-art classification performance for the binary classification task. The recognition accuracies reach to 96.16% and 95.89% for valence and arousal. The proposed method can improve 3% to 6% than other latest advanced works.