The Study Of Classification Method Based On EEG In Visual Search

Visual search is a critical cognitive function to quickly identify target objects in visual scenes.The covert shifts of attention in visual search processing produces a lateral N2 pc event-related potential,which is recorded at the posterior electrode about 200 ms after the search display onset.According to this characteristic,the N2 pc can be used to determine the general direction of the lateral target in a visual search task,which has great significance to brain-computer-interface application requiring the covert orienting of attention.Electroencephalograph was the most used signal acquisition method in brain-computer interface control because of the high temporal resolution and non-invasive measurement.This paper explores the EEG-based visual search classification method to identify a singleton target.This paper includes the following three aspects:First,this study investigated the visual search classification method with N2 pc amplitude features.Classification of the N2 pc could be used to detect the attentional shifts associated with lateral targets and predict the spatial position of the lateral target because of its lateralized scalp distribution in visual search.The multiple correlated component analysis(MCORCA)method was developed to identify lateral targets in a serial search task with single trial and it can extract the linear combination of optimal correlation components.The average classification accuracy at the scalp and cortical levels reached 82% and 84%,respectively.These results showed that the MCORCA-based classification method could improve the classification performance of N2pc-based brain-computer interface.The N2 pc could be used for visual search classification.Second,this study investigated the visual search classification method based on brain networks.Weighted minimum norm estimates was used to reconstruct the source time series that underlie scalp EEG data.The wavelet coherence analysis was used to evaluate the functional connectivity across brain regions during the visual search processing.The difference network between the left visual field and right visual field was constructed,and then the left and right visual field targets were identified based on the wavelet coherence values for the difference network connections with support vector machine classifier.The results revealed that the theta rhythm of at 200-400 ms after stimulus onset achieved the best classification performance,achieving an average classification accuracy of 87% in serial search task.It shows that N2 pc is closely related to theta rhythm,and the wavelet coherence network features in theta rhythm could be used for visual search classification.Third,this study investigated the visual search classification method based on deep transfer learning.The transfer learning with deep convolutional neural network for visual search classification was explored.The EEG data recorded in experimental scene were transferred to the real visual search task via a convolutional network.A spatio-temporal convolutional neural network(STCNN)with a first convolution across time and a second convolution across space(electrodes)was constructed.The cross-subject and cross-experiment data were used to train the source domain STCNN.Next,the target domain STCNN model was trained by transferring the STCNN model from source domain to target domain and performing fine-tuning in target domain.The classification results achieved an average classification accuracy of 86.1% in real visual search scenes,indicating the potential of the deep transfer learning method for visual search classification.