Swin-CANet: A Novel Integration of Swin Transformer with Channel Attention for Enhanced Motor Imagery Classification
Yufan Shi, Menghan Wang · 2024
Electroencephalogram (EEG) signals are rich in physiological and psychological information, and its decoding technology enables machines to recognize neural activities, which is significant for research in fields such as Brain Computer Interfaces (BCIs) and medical rehabilitation. Given that the temporal dimension of EEG signals far exceeds its channel dimension, researchers generally tend to explore the temporal context information of EEG signals to capture the subtle differences in Motor Imagery (MI), while relatively ignoring the importance of different electrode channels. In fact, cross-channel information contains an intrinsic semantic structure that is indispensable for improving the classification performance of EEG signals. In the feature extraction stage, enhancing the weight of the channels related to motor imagery while reducing the weight of unrelated channels can facilitate the extraction of spatial semantic features in EEG signals, thus improving the classification accuracy. The key to deeper decoding of the brain's cognitive functions is how to synchronously parse and represent the spatial and temporal dimensional properties of neural activities with high temporal resolution. In this paper, we innovatively propose a Swin Transformer with Channel Attention Network model, Swin-CANet, which aims to extract and construct superior, robust, and discriminative spatiotemporal feature representations from raw EEG signals and reveal the deep-level correlations between spatiotemporal features. Further, we design a lightweight module, Swin-CA Block, effectively reducing the computational complexity. We prove the effectiveness and superiority of the Swin-CANet model on the BCI IV-2a dataset. The experimental results show that the classification accuracy of Swin-CAN et is improved by 4.28% to 11.03% compared with the existing five models.