Research On EEG Emotion Recognition Based On GoogLeNet

Emotion is a complex mental state,which affects human physiological and psychological behaviors at all times,and accurate recognition of emotions has practical application prospects in the medical field.Owing to the characteristics of high temporal resolution,strong emotional correlation and wide spatial distribution of EEG,EEG emotion recognition has gradually become a popular research topic in recent years.In the field of EEG emotion recognition,machine learning methods and deep learning methods are mainly used.The former is gradually eliminated due to the defects of low recognition accuracy and long computation time,while the latter is gradually popularized due to high recognition accuracy,short computation time and strong fitting ability.Considering the advantages and disadvantages of these two methods,we proposes an efficient EEG signal data preprocessing algorithm that saves time and reduces computation,followed by an EEG emotion recognition model that can extract and fuse dimensional features in time domain,spatial domain and global based on the GoogLeNet model,and finally uses an attention mechanism for our model so that the baseline model achieves high accuracy and fast inference.The main research and innovation points within this paper are as follows.(1)In the paper,four pre-processing methods will be used for DEAP data set by combining the temporal sequence of EEG itself and the spatial domain of brain electrode placement,and the better EEG signal pre-processing methods will be selected through relevant experiments.The experimental results show that the Z-Score preprocessing method is the best for the EEG data channel(brain electrode-spatial domain),and the preprocessing method can effectively reduce the model computation and improve the accuracy of the model for emotion recognition.(2)We proposes a convolutional neural network model for EEG emotion recognition based on the GoogLeNet.The model uses time-domain convolutional kernels,spatial-domain convolutional kernels and global convolutional kernels in stages.By using these convolutional kernels,the model can extract emotion-related features in multiple dimensions and finally use them for emotion classification.The experimental results show that the accuracy of the GTSCeption model is 8.76% higher than the current best Inception-based emotion recognition model in the Arousal emotion dimension,and the accuracy of the Valence emotion dimension can reach up to 94.15%.(3)In the paper,an attention mechanism is introduced to the baseline model proposed in(2).The attention mechanism is used to explore the important emotional information of EEG signals in the time and frequency domains,thus improving the accuracy of emotion recognition.It is experimentally demonstrated that the baseline model based on the SE attention mechanism has the most significant improvement in accuracy.