Zhang Guokai, Luo Jihao, Han Letong, Lu Zhuyin, Hua Rong, Chen Jianqing, Che Wenliang
School of Optical-Electrical and Computer Engineering, University of Shanghai for Science and Technology, Shanghai, China.
School of Software Engineering, Tongji University, Shanghai, China.
Front Neurosci. 2021 Jan 13;14:578255. doi: 10.3389/fnins.2020.578255. eCollection 2020.
Accurate and automatic classification of the speech imagery electroencephalography (EEG) signals from a Brain-Computer Interface (BCI) system is highly demanded in clinical diagnosis. The key factor in designing an automatic classification system is to extract essential features from the original input; though many methods have achieved great success in this domain, they may fail to process the multi-scale representations from different receptive fields and thus hinder the model from achieving a higher performance. To address this challenge, in this paper, we propose a novel dynamic multi-scale network to achieve the EEG signal classification. The whole classification network is based on ResNet, and the input signal first encodes the features by the Short-time Fourier Transform (STFT); then, to further improve the multi-scale feature extraction ability, we incorporate a dynamic multi-scale (DMS) layer, which allows the network to learn multi-scale features from different receptive fields at a more granular level. To validate the effectiveness of our designed network, we conduct extensive experiments on public dataset III of BCI competition II, and the experimental results demonstrate that our proposed dynamic multi-scale network could achieve promising classification performance in this task.
在临床诊断中，对脑机接口（BCI）系统的语音想象脑电图（EEG）信号进行准确且自动的分类有很高的要求。设计自动分类系统的关键因素是从原始输入中提取基本特征；尽管许多方法在该领域取得了巨大成功，但它们可能无法处理来自不同感受野的多尺度表示，从而阻碍模型实现更高的性能。为应对这一挑战，在本文中，我们提出了一种新颖的动态多尺度网络来实现EEG信号分类。整个分类网络基于ResNet，输入信号首先通过短时傅里叶变换（STFT）对特征进行编码；然后，为进一步提高多尺度特征提取能力，我们引入了一个动态多尺度（DMS）层，该层允许网络在更精细的层面上从不同感受野学习多尺度特征。为验证我们设计的网络的有效性，我们在BCI竞赛II的公共数据集III上进行了广泛的实验，实验结果表明，我们提出的动态多尺度网络在该任务中可以实现有前景的分类性能。
