EEG-Based Emotion Recognition Using Temporal Convolutional Network

Emotion recognition based on physiological signal can be used in many applications such as, intelligent human-computer interface design, emotional disorder diagnoses. For traditional approaches, the prior knowledge is required to design and extract a range of features from physiological signal. The generalization ability of traditional methods is poor because of the lack of high-level features. Using deep-learning methodologies to analyze physiological signal, i.e. eeg, becomes increasingly attractive for recognizing emotions. In this paper, we design a sequence model based on deep-learning that uses Temporal Convolutional Network(TCN) to extract high-level features in consideration of the time dependence of physiological signals for EEG emotion recognition. Specifically, we extract the differential entropy feature in seconds and construct a time series with fixed-length time window data as the input to TCN, and then using softmax to classify. Furthermore, in order to get reliable results, we divide the samples according to the trials, avoiding the testing set samples and training set samples from the same trial. Specifically, we first divide the samples according to the trials as the testing set and the training set, and then segment the trials in the testing set and training set with fixed time window length to obtain more samples respectively. To evaluate the performance of the proposed model, we conduct the emotion classification experiments on DEAP database. The experimental results show the effectiveness of our proposed model for EEG emotion recognition.