Motor Imagery based EEG Signal Classification using Multi-scale CNN Architecture

Deep learning has been employed in EEG signal classification in recent years since it has multiple advantages, including automatic feature extraction and classification. Although deep learning models have been used recently for motor imagery signal classification, still there are challenges to improve the performance of classification. One of the challenges is the small size of the training dataset. Also, most of the existing CNN models utilize the features extracted from the last layer, which might result in the loss of some specific details. A multi-scale CNN architecture is proposed in this work to utilize features at the initial and final layers for classification of MI signals. Firstly, the size of the dataset is increased by using the sliding window data augmentation technique and time-frequency images are obtained after applying short time Fourier transform (STFT). The images are then fed to the multi-scale CNN. A mean kappa value of 0.803 and a mean classification accuracy of 85.3 percent is achieved for this model on dataset 2a of BCI competition IV through subject-specific training.