Visual EEG Classification Based On Attention Mechanism

Brain-Computer Interface(BCI)is a technology that enables human-computer interaction by recording and analyzing brain signals.Visual classification refers to the task of recognizing objects using human visual perception.The use of BCI technology to achieve visual classification tasks is referred to as visual EEG classification.Participants observe images of different object categories and use deep neural networks to learn brain signals induced by visual stimuli to distinguish between multiple visual object categories and classify the object images.The brain electrical signals contain information of temporal,spectral,and spatial dimensions.Existing visual brain-computer interface(BCI)methods mainly extract features in the time or frequency domain.Although these methods have made certain progress,there are still deficiencies that need to be addressed to further improve classification accuracy.As the subjects usually perceive local information of objects at different time points,obtaining global information of brain signals is crucial for the classification task that requires considering the overall structure of objects.Existing models mostly rely on the time dimension and use convolution or gating mechanisms to extract partial features,making it difficult to obtain the global features of brain signals.To address these issues,this thesis introduces the attention mechanism and constructs a visual BCI model in different dimensions of time,frequency,and space,which can better enhance the saliency of signal features,thus improving the accuracy of signal processing and classification.The main contributions of this thesis are summarized as follows:To address the problem of perceiving global features in the temporal dimension of visual EEG signals,this thesis introduces attention mechanisms to better connect contextual information,and proposes a bidirectional long short-term memory visual classification model based on global attention mechanisms and a Transformer visual classification model based on time series,to better obtain and perceive global features.The classification accuracies are 94.15% and 95.27%,respectively.To address the problem of low temporal salience in visual EEG classification tasks,this thesis proposes a visual classification model based on the fusion of time-frequency features.The frequency domain features also contain global information of visual EEG signals.The time and frequency domain features are applied in parallel in the Transformer module,and are fused by concatenation.This method can better improve signal salience,and achieves a classification accuracy of98.99%.To address the lack of different dimensional feature fusion methods in visual EEG classification tasks,this thesis proposes a visual classification model based on the fusion of time-frequency-space features.EEG signal channels contain spatial topological structure information of the brain perception area.Spatial attention modules are used to extract feature information of EEG signals in the spatial dimension.Various methods of time-frequency-space feature fusion are discussed by combining time-frequency features.The classification accuracy reaches 99.39%.Finally,the model is applied to other types of cognitive EEG tasks,verifying the superiority and robustness of the fusion method.In this thesis,by constructing a visual EEG classification model based on attention mechanism,we investigate the way of feature fusion in different dimensions of time,frequency and space,so as to demonstrate its application value and potential prospect in BCI technology.