Enhancing EEG Signals Classification through Spatio-Temporal Feature Fusion in Dense Connection Neural Networks

Brain-computer interface (BCI) technology based on electroencephalogram (EEG) has attracted significant interest, particularly in EEG signal interpretation, pattern recognition, and brain activity classification, which are considered promising research avenues. Nonetheless, the classification of target signals using EEG continues to pose substantial challenges regarding the performance and interpretability of human brain signals. To address these issues, this study proposes a novel dense connection convolutional neural network with spatio-temporal feature fusion to classify brain visual images. Drawing inspiration from visual attention and brain memory mechanisms, a densely connected module is incorporated to mitigate the vanishing gradient issue and facilitate effective training of deep networks. EEG signals are subsequently encoded and stored along both temporal and spatial dimensions. To leveraging the unique properties of EEG signals, a bidirectional gated recurrent unit network is employed to derive temporal features, while spatial features are extracted using a two-dimensional mixed dilated convolution module. Ultimately, the extracted spatio-temporal features are concatenated and classified. The findings confirm the feasibility and effectiveness of the proposed model.