Performance Evaluation of Xception Networks and Short-Time Fourier Transform Spectrograms for Motor Imagery Classification

This study seeks to explore the success of the Xception convolutional neural network and short-time Fourier transform in classifying EEG signals with limited training samples using pre-trained deep transfer learning. So far, most deep learning research has concentrated on one-dimensional time series inputs. Two-dimensional inputs are another excellent way of giving EEG signals to deep learning. This study used a two-dimensional STFT-based method to transform EEG signals into images and Xception for classification. BCI Competition IV dataset 2b, which includes nine participants, was used to evaluate the performance. This study is the first to report the STFT + Xception approach results for classifying motor imagery signals. It achieved good results from a small number of MIEEG data, averaging over 80% for nine subjects with a very low standard deviation of 2.9% between subjects. Applying data augmentation and training from scratch with more data can show more successful results in the future.