In recent years, there has been a significant increase in research activity on electroencephalography (EEG)-based motor imagery brain-computer interfaces (MI-BCI) in the field of deep learning. However, despite achieving high accuracy, the size of models is increasing, requiring significant memory and computational resources. Therefore, finding a balance between accuracy and computational cost has always been a challenge in MI classification research. Convolutional Neural Networks (CNNs) generate feature representations of objects by collecting semantic sub-features. The activation of subfeatures is susceptible to noisy backgrounds. The Spatial Group-wise Enhance (SGE) module adjusts the importance of each sub-feature by generating an attention factor for the spatial location of each semantic group, thus enhancing useful features and suppressing noise. The design of the SGE module is lightweight, with few parameters and computations. Therefore, we introduce the SGE module to improve accuracy and minimize model parameters. In this paper, we propose EEG-SGENet, a novel end-to-end convolutional neural network model that considers both the lightweight model and accuracy. Experimental results on the BCI IV 2a dataset show that EEG-SGENet achieves an accuracy of 80.98% in the four categories of MI. The average classification accuracy for the two-category task of BCI IV 2b is 76.17%. Comparisons with other lightweight models in terms of classification accuracy and other aspects have shown that this model achieves a good balance between decoding performance and computational cost. Overall, experimental results demonstrate that the proposed model is expected to become a new method for decoding EEG signals.
