Emotion recognition from EEG using RASM and LSTM

Abstract

In the field of human-computer interaction, automatic emotion recognition is an important and challenging task. As a physiological signal that directly reflects the brain activity, EEG has advantages in emotion recognition. However, previous studies seldom consider together the temporal, spatial, and frequency characteristics of EEG signals, and the reported emotion recognition accuracy is not adequate for applications. To address this issue, this study proposes a new approach which extracts RASM as the feature to describe the frequency-space domain characteristics of EEG signals and constructs a LSTM network as the classifier to explore the temporal correlations of EEG signals. It is implemented on the DEAP dataset for a trial-level emotion recognition task. In a comparison with a number of relevant studies on DEAP, its mean accuracy of 76.67% ranks the first, which approves the effectiveness of this new approach.