A Transfer Learning Method based on VGG-16 Convolutional Neural Network for MI Classification

Brain computer interface (BCI) technology can help the disabled to achieve the recovery of neural function by using the Motor Imagery Electroencephalogram (MI-EEG) based rehebilitation system. However, it is difficult to acquire a large amount of available EEG data, transfer learning technology provides an effective method, and the source domain selection is one of key issues. In this study, we develop a novel parameter transfer learning method based on VGG-16 convolutional neural network (CNN) for MI classification. First, the number of fall MI-EEG signals are augmented with the sliding window method, and the short-time Fourier transformation (STFT) is applied to obtain the time-frequency spectrum images (TFSI). Then, the VGG-16 CNN is pre-trained with TFSI of source domain, which is divided into five blocks.. The parameters of the pre-trained CNN are transferred to the target network though a new transfer strategy, i.e. utilization of the data of part subjects from target domain to fine-tune the five blocks in turn. Finally, the fine-tuned CNN is used for MI classification of the rest subjects in target domain. This work is evaluated with a public dataset, the best classification accuracy of this study is 96.59%. The results show that the high correlation between the source domain and the target domain is better than using the domains with low correlation, and the proposed transfer strategy is efficiency.

Automated summary

The study develops a novel MI classification method based on VGG-16 convolutional neural network using transfer learning technology to aid disabled individuals in neural function recovery. By pre-training and fine-tuning the CNN, a high classification accuracy of 96.59% is achieved, demonstrating the effectiveness of utilizing high correlation between source and target domains.