Classification of EEG Motor Imagery Tasks Using Convolution Neural Networks

Electroencephalograph (EEC) is a highly nonlinear data and very difficult to be classified. The EEC signal is commonly used in the area of Brain-Computer Interface (BCI). The signal can be used as an operative command for directional movements for a powered wheelchair to assist people with disability in performing the daily activity. In this paper, we aim to classify Electroencephalograph EEC signals extracted from subjects which had been trained to perform four Motoric Imagery (MI) tasks for two classes. The classification will be processed via a Convolutional Neural Network (CNN) utilising all 22 electrodes based on 10-20 system placement. The EEC datasets will be transformed into scaleogram using Continuous Wavelet Transform (CWT) method. We evaluated two different types of image configuration, i.e. layered and stacked input datasets. Our procedure starts from denoising the EEG signals, employing Bump CWT from 8-32 Hz brain wave. Our CNN architecture is based on the Visual Geometry Croup (VCC-16) network. Our results show that layered image dataset yields a high accuracy with an average of 68.33% for two classes classification.