Motor Imagery Tasks EEG Signals Classification Using ResNet with Multi-Time-Frequency Representation

This paper presents a ResNet-based multi-feature fusion method for motor imagery task classification in brain-computer interfaces. To address the drawbacks of insufficient information and low signal-to-noise ratio extracted from EEG signal features by a single time-frequency transform. Three time-frequency transforms including morlet wavelet transform, multi-window time-frequency transform and stockwell transform are used to extract features from EEG signals. Those features are then input to three ResNets for further feature extraction, and finally the high-level feature outputs by the three networks are fused before the fully connected layer of the original ResNet, and then fed into a fully connected layer for classification. The performance of the proposed method is evaluated using the accuracy and kappa value on the IVa dataset of BCI Competition III, and the results show that the ResNet model using feature fusion produces better results with a classification accuracy of 97.86%, which is 39.65% higher than the model using only one time-frequency transform as input.