DBANet: a dual-branch convolutional neural network with attention enhancement for motor imagery classification

Background Recently, Motor Imagery (MI) Electroencephalography (EEG) decoding has been extensively applied in diverse brain–computer interface (BCI) systems. Deep learning is widely used for signal processing in MI-BCI due to the reliability of automatic extracting features. However, recent deep learning methods fail to capture the multi-dimensional interaction of features effectively.New method This research proposes a Dual-Branch Convolutional Neural Network with Attention Enhancement (DBANet) for decoding MI. Firstly, we use a filter bank alignment module, it aligns the multi-frequency data and reduce the differences in the MI data. Subsequently, a spatial-temporal module to extract the spatial-temporal features is employed. Then, a transpose attention module is integrated to extract distinct features from various brain regions and temporal windows. Finally, the combined features are applied for classification.Results The subject-dependent results of our proposed framework on the three datasets (BCI Competition IV dataset 2b, 2a and ECUST dataset) are 85.19%, 85.15% and 75.24%, respectively.Comparison with existing methods We conduct an extensive study between the proposed framework and five State-of-the-Art models, including FBCSP, ShallowConvNet, EEGNet, FBCNet, and IFNet.Conclusions for research articles We certificate the superiority of the proposed framework by conducting comparative experiments on three datasets with advanced MI decoding algorithms.