Electroencephalogram Emotion Recognition Based on Manifold Geomorphological Features in Riemannian Space

Most of the electroencephalogram (EEG) emotion recognitions are conducted in linear euclidean space. However, it is difficult to accurately describe the nonlinear characteristics of multivariate EEG signals. Comparatively, the Riemannian manifold is a nonlinear space in which features of multivariate EEGs can be analyzed more thoroughly. Therefore, inspired by geographical knowledge, an EEG emotion recognition methodology based on geomorphological features of the Riemannian manifold (GFRM) is proposed. First, in terms of the Wasserstein scalar curvature, an automatic search strategy is developed to narrow down the domain of interest so as to reduce the computation load. Afterward, the geomorphological homogeneity function (GHF) is designed to evaluate regional features of the Riemannian manifold. Finally, we simultaneously devised the fuzzy $\mathbf{k}$k-nearest neighbor classifier of the Riemannian manifold and the local mean classifier of the Riemannian manifold for recognition. On the basis of the GHF, GFRM can automatically choose an appropriate classification strategy for every specific instance to greatly raise the efficiency and accuracy. Two public datasets and one practical lab dataset are utilized to validate the performance of GFRM.