EEG-Based Motor Imagery Classification via Multi-scale Convolutional Network and Improved Capsule Network

Abstract Electroencephalogram (EEG) based motor imagery classification is a crucial component of brain-computer interfaces (BCIs). Traditionally, convolutional neural networks (CNNs) have been extensively employed for this task. However, due to local feature learning mechanism, CNNs have difficulty in capturing the global contextual information from EEG signal, which limit the performance of brain signal decoding. In this study, in order to overcome the shortcomings of CNNs, we introduce an improved capsule network to effectively learn various properties within EEG features and characterize intrinsic relationships between EEG features, achieving more robust performance. A novel end-to-end model is proposed in this paper which integrates muti-scale convolutional network and improved capsule network with self-attention routing mechanism, namely MSC-CapsNet. In the proposed model, a multi-scale convolutional network is employed to fully learn spatial and temporal information from EEG signals and encode them into discriminative features, then a improved capsule network with self-attention routing mechanism is applied to convert EEG features into entities corresponding to motor imagery classes and output classification results. The proposed model achieves the state-of-the-art performance on public Competition 4-2a dataset without using any data augmentation operations, with average accuracy of 86. 1\%. This study demonstrates the great potential of capsule network in EEG decoding and is expected to become a general architecture to improve robustness and generalization capabilities.