Data-driven Data Augmentation for Motor Imagery Brain-Computer Interface

When utilizing deep learning (DL) for the braincomputer interface (BCI), it suffers from degradation of classification performance due to the small quantity of electroencephalography (EEG) training data. Typically, artifacts such as blinking and a falling-off in user’s attention caused by a long experiment period have a great effect on the small quantity and the low quality of the EEG training data. To address this problem, in this work, we introduce a novel classification method based on data augmentation method and DL framework for motor imagery (MI) EEGs classification. The proposed framework generates input EEG images with distinct MI features from a small amount of EEG training data. Experimental results using a public BCI dataset show that the proposed method outperforms the predecessors which use a simple DL model.