Enhancing Performance of EEG-based Emotion Recognition Systems Using Feature Smoothing

Electroencephalography (EEG) has been used recently in emotion recognition. However, the drawback of current EEG-based emotion recognition systems is that the correlation between EEG and emotion characteristics is not taken into account. There are the differences among EEG features, even with the same emotion state in adjacent time because EEG extracted features usually change dramatically, while emotion states vary gradually or smoothly. In addition, EEG signals are very weak and subject to contamination from many artefact signals, thus leading to an accuracy reduction of emotion recognition systems. In this paper, we study on feature smoothing on EEG-based Emotion Recognition Model to overcome those disadvantages. The proposed methodology was examined on two useful kinds of features: power spectral density (PSD) and autoregressive (AR) for two-level class and three-level class using DEAP database. Our experimental results showed that feature smoothing affects to both the feature sets, and increases the emotion recognition accuracy. The highest accuracies are 77.38 % for two-level classes and 71.75 % for three-level classes, respectively in valence space.