CNN models in time-frequency domain for identification of motor imagery tasks from EEG signals.

The Motor imagery (MI) based brain computer interface (BCI) system provides a way for the people suffering from motor impairments to communicate to the external world. This work proposes compact convolutional neural network (CNN) models with single convolutional layer, for effectively classifying the left and right hand MI tasks using the EEG data from only two channels. The Complex Morlet Wavelets (CMW) are used here to extract high-resolution time and frequency domain features from MI EEG signal. These time-frequency representations (TFR) serve as inputs to three proposed CNN models namely the time domain CNN (TD-CNN), the frequency domain CNN (FD-CNN) and the time-frequency domain CNN (TF-CNN) models, which perform a convolution along the time, frequency and time-frequency domain features of the data respectively. The developed models have been evaluated on the BCI Competition 4 dataset 2a using the subject-dependent and subject-independent validation strategies. The TF-CNN model has outperformed the TD-CNN and FD-CNN models, by giving a classification accuracy of 85.83% and 77.6% for the subject dependent and independent validations respectively. The results show that the proposed models have given a better performance than the state-of-the art methods and the existing CNN models with complex network architectures.