Kant Piyush, Laskar Shahedul Haque, Hazarika Jupitara, Mahamune Rupesh
Department of Electronics and Instrumentation Engineering, National Institute of Technology, Silchar 788010, Assam, India.
Department of Electronics and Instrumentation Engineering, National Institute of Technology, Silchar 788010, Assam, India.
J Neurosci Methods. 2020 Nov 1;345:108886. doi: 10.1016/j.jneumeth.2020.108886. Epub 2020 Jul 28.
The processing of brain signals for Motor imagery (MI) classification to have better accuracy is a key issue in the Brain-Computer Interface (BCI). While conventional methods like Artificial neural network (ANN), Linear discernment analysis (LDA), K-Nearest Neighbor (KNN), Support vector machine (SVM), etc. have made significant progress in terms of classification accuracy, deep transfer learning-based systems have shown the potential to outperform them. BCI can play a vital role in enabling communication with the external world for persons with motor disabilities.
Deep learning has been a success in many fields. However, for Electroencephalogram (EEG) signals, relatively minimal work has been carried out using deep learning. This paper proposes a combination of Continuous Wavelet Transform (CWT) along with deep learning-based transfer learning to solve the problem. CWT transforms one dimensional EEG signals into two-dimensional time-frequency-amplitude representation enabling us to exploit available deep networks through transfer learning.
The effectiveness of the proposed approach is evaluated in this study using an openly available BCI competition data-set. The results of the approach have been compared to earlier works on the same dataset, and a promising validation accuracy of 95.71% is achieved in our investigation.
Our approach has shown significant improvement over other studies, which is 5.71% improvement over earlier reported algorithm (Tabar and Halici, 2017) using the same dataset. Results show the validity of the proposed Deep Transfer-Learning based technique as a state of the art technique for MI classification in BCI.
为提高运动想象（MI）分类的脑信号处理精度是脑机接口（BCI）中的一个关键问题。虽然传统方法如人工神经网络（ANN）、线性判别分析（LDA）、K近邻（KNN）、支持向量机（SVM）等在分类精度方面取得了显著进展，但基于深度迁移学习的系统已显示出超越它们的潜力。BCI在使运动障碍者与外部世界进行通信方面可以发挥至关重要的作用。
深度学习在许多领域都取得了成功。然而，对于脑电图（EEG）信号，使用深度学习进行的工作相对较少。本文提出将连续小波变换（CWT）与基于深度学习的迁移学习相结合来解决该问题。CWT将一维EEG信号转换为二维时频幅度表示，使我们能够通过迁移学习利用现有的深度网络。
本研究使用公开可用的BCI竞赛数据集评估了所提出方法的有效性。该方法的结果已与同一数据集上的早期工作进行了比较，在我们的研究中实现了95.71%的有前景的验证精度。
我们的方法相对于其他研究有显著改进，与使用相同数据集的早期报道算法（Tabar和Halici，2017）相比提高了5.71%。结果表明所提出的基于深度迁移学习的技术作为BCI中MI分类的一种先进技术是有效的。
