A lightweight and accurate double-branch neural network for four-class motor imagery classification

Deep learning is an important pathway for investigation of motor imagery signal classification. Nevertheless, maintaining a good compromise between performance and computational cost has been a major challenge in developing deep models for decoding motor imagery EEG. In this paper, a novel shallow double-branch con-volutional neural network (DSCNN) is proposed for four-class motor imagery classification. The proposed CNN adopts parallel extraction of two branches to improve classification accuracy. Meanwhile, in order to constrain the depth of the whole network, the left branch only contained two single temporal and spatial convolutional layers to extract common EEG features. Similarly, the right branch first introduced 1D convolution to exploit the channel dependency and temporal features across multiple time-scales, secondly a depth-wise separable con-volutional layer was applied for optimizing EEG signal series. Then the feature representation for final classi-fication was obtained by merging intermediate features extracted from the two branches. Also, the DSCNN is an end-to-end decoder, as it employs the raw EEG data as inputs and does not require additional complex pre -processing. The proposed model was evaluated on public benchmark BCI competition IV dataset 2a and achieved in terms of accuracy is 85% and kappa value is 0.79. Compared with other state-of-the-art algorithms, the experiment results reveal that the DSCNN has higher decoding accuracy and robustness, as well as a 10% improvement in accuracy than the single general shallow model. Furthermore, as a lightweight architecture, the DSCNN relies on a lower computational power than similar mainstream models, which is more in line with the requirements of low delay and real-time performance in practical BCI applications.

Automated summary

This paper proposes a novel shallow double-branch convolutional neural network (DSCNN) for decoding motor imagery EEG. By adopting parallel extraction and merging intermediate features, the proposed DSCNN achieves higher decoding accuracy and robustness compared to other algorithms, while requiring lower computational power.