Electroencephalography (EEG) has relatively poor spatial resolution and may yield incorrect brain dynamics and distort topography; thus, high-density EEG systems are necessary for better analysis. Conventional methods have been proposed to solve these problems, however, they depend on parameters or brain models that are not simple to address. Therefore, new approaches are necessary to enhance EEG spatial resolution while maintaining its data properties. In this work, we investigated the super-resolution (SR) technique using deep convolutional neural networks (CNN) with simulated EEG data with white Gaussian and real brain noises, and experimental EEG data obtained during an auditory evoked potential task. SR EEG simulated data with white Gaussian noise or brain noise demonstrated a lower mean squared error and higher correlations with sensor information, and detected sources even more clearly than did low resolution (LR) EEG. In addition, experimental SR data also demonstrated far smaller errors for N1 and P2 components, and yielded reasonable localized sources, while LR data did not. We verified our proposed approach's feasibility and efficacy, and conclude that it may be possible to explore various brain dynamics even with a small number of sensors.
Keywords: convolutional neural networks; electroencephalography; spatial resolution; super-resolution.