EEG-Based Classification of Lower Limb Motor Imagery with STFT and CNN

• Published in: Neural Information Processing Vol. 1517; pp. 397 - 404
• Main Authors: Lu, Boyang, Ge, Sheng, Wang, Haixian
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer International Publishing AG 2021; Springer International Publishing
• Series: Communications in Computer and Information Science
• Subjects: Convolutional neural network; Lower limb; Short-time Fourier transform
• DOI: 10.1007/978-3-030-92310-5_46
• ISBN: 9783030923099; 3030923096
• ISSN: 1865-0929

In order to classify the brain signals of lower limb motor imagery, we used the method of short-time fourier transform (STFT) to transform the signals into time spectrum, and then processed the size and gray scale of the obtained time spectrum. Thus we constructed a neural network model called pragmatic convolutional neural network (pCNN), and the processed 128 * 128 pixel grayscale time spectrums were used as the input for classification. The classification effect was good on all 10 subjects, with the highest accuracy reaching 76% $$\%$$ , while the comparison model was only 66.88% $$\%$$ (shallow CNN), 52% $$\%$$ (recurrent CNN) and 68.62 (common spatial pattern + support vector machines). The research results show that STFT is very effective in transforming the EEG input of CNN, and due to the difference of the activated regions between lower limbs and upper limbs, many models with good performance for upper limbs cannot be simply copied to lower limbs.