Robust EEG-based Emotion Recognition using Multi-feature Joint Sparse Representation

Abstract

The key problem of emotion recognition lie in effective electroencephalography (EEG) signal processing and feature extraction. Traditional methods mostly extract single feature or simply combine different features together, but emotion recognition based merely on a single feature may be unreliable and simple feature combination will be affected by complex feature interdependencies. To improve recognition accuracy, we propose a multi-feature emotion recognition method based on Joint Sparse Representation (JSR) to transform the simple feature fusion into an optimization problem. Specifically, sparse matrices for each individual feature are combined to obtain the JSR of these features, and three different EEG features including wavelet energy, Hurst index, and fractal dimension are employed to produce multi-feature fusion results. Due to high computational complexity and restrictions on kernel function selection of Support Vector Machine (SVM), we adopt Relevance Vector Machine (RVM) to classify emotions. Simulation results show our proposed multi-feature fusion algorithm has an average recognition accuracy of over 85%, which is 8% higher than the traditional method.