A Novel Signal-to-Image Conversion Approach with Ensembles of Pretrained CNNs for Motor Imagery EEG Signals

Deep learning techniques can recognize and learn significant features automatically. The advantages of deep learning should also be used to solve problems in other fields. Therefore, we concentrated on classifying motor imagery EEG signals with deep transfer learning. The present study first proposed an EEG signal-to-image conversion method that uses 2D signal representations. Thanks to this, we employed 2D images as input instead of standard one-dimensional EEG features. Second, we used AlexNet, GoogLeNet, and SqueezeNet pre-trained convolutional neural network frameworks for classification. We also used hard voting and merged the outputs of pre-trained convolutional neural networks. We performed experiments on dataset 2a of BCI Competition IV, achieving 84.18±5.37% classification accuracy and 0.66±0.12 kappa across nine subjects, even with limited training data. The results showed that the proposed signal-to-image conversion method with deep transfer learning has potential, particularly in processing biomedical signals.