Motor imagery electroencephalogram (EEG)-based brain-computer interfaces (BCIs) aid individuals with restricted limb mobility. However, challenges like low signal-to-noise ratio and limited spatial resolution hinder accurate feature extraction from EEG signals, impacting classification. To tackle these issues, we propose an end-to-end dual-branch neural network (EEG-DBNet). This network decodes temporal and spectral sequences separately using distinct branches. Each branch has local and global convolution blocks for extracting local and global features. The temporal branch employs three convolutional layers with smaller kernels, fewer channels, and average pooling, while the spectral branch uses larger kernels, more channels, and max pooling. Global convolution blocks then extract comprehensive features. Outputs from both branches are concatenated and fed to fully connected layers for classification. Ablation experiments demonstrate that our architecture, with specialized convolutional parameters for temporal and spectral sequences, significantly improves classification accuracy compared to single-branch structures. The complementary relationship between local and global convolutional blocks compensates for traditional CNNs' limitations in global feature extraction. Our method achieves accuracies of 85.84% and 91.42% on BCI Competition 4-2a and 4-2b datasets, respectively, surpassing existing state-of-the-art models. Source code is available at https://github.com/xicheng105/EEG-DBNet.
View on arXiv