Feasibility Study of EEG Super-Resolution Using Deep Convolutional Networks

The success of deep learning for super-resolution (SR) in image processing motivated us to investigate whether it is feasible and applicable to electroencephalography (EEG) data. We considered three questions: 1) How does noise type (white Gaussian or colored) and its signal-to-noise ratio (SNR) affect the EEG SR process? 2) How does SR work over various upscaling sizes? 3) Are there any approaches to improve signal quality when we perform SR? In this work, we proposed deep convolutional networks to enhance the spatial resolution of simulated EEG data. In the simulation of white Gaussian noise, we observed that the SR not only altered the signal from low-resolution (LR) to high-resolution (HR), but also improved signal quality. In the real (colored) noise, it recovered the signal to the level of its target data. Even when the upscaling ratio of SR increased, the signal quality obtained was acceptable. The limitation in reproducing real noisy EEG data may be overcome by applying whitening technique. It is expected that EEG SR can reduce experimental costs significantly, thus is quite promising.