Dual-frequency SSVEP Based Color And Orientation Visual Feature Extraction

The research in pattern classification of brain physiological signals,including EEG\MEG signals and fMRI BOLD signals,has always been in research focus in the field of cognitive neural science,not only for the academic,but also for the practical application.By the application of multi-variable pattern alalysis in brain physiological analisis,we can reveal more information that can’t be easily accessed by regular researching methods.Except for classification,the calulation of linear model is also a more and more commonly uesd method.More and more related research were trying to mathematically model the visual cortex especially the V1 according to the known research findings,and use the model to reconstruct the features in the visual stimulation.The experiment in this paper is mainly about EEG signal.The objective of our research was to separately extract the signal features of two simultaneously.presented visual stimulation and use pattern analysis.The major works involved in this paper and the correspongding results are listed as below:1)The experimental paradigm design of dual-frequency flicking visual stimulation presentation for steady-state visual evoked potential(SSVEP).We used two frequencies to present two kinds of visual stimuli color and orientation,including 6 colors and 6 orientations,respectively,and each is presented by the manner of “contrast reversal pattern”.There were attend-to-color and attend-to-orientation two conditions in the experiment.2)We calculated the forward encoding model for each visual stimuli presented in the experiment,and a significant effect was fount between attended and unattended condition.The forward encoding model was also used to adjust the signal feature parameters to get the best signal to noise ration feature.3)In addition to using the power spectrum feature,two kinds of signal feature extraction methods are proposed.First kind of feature is extracted from time-frequency map by the mechanism of SSVEP power amplitude effected by visual attention.The second is using forward encoding model to extract the optimal SNR(signal to noise ratio)features.4)Totally three kinds of signal feature and three kinds of classifier,including linear discriminant analysis,stacked auto-encoder neural network and support vector machine,were used to performe the data analysis.A six-category classification within 6 different colors and within 6 orientations were performed,then a binary classification between attend-to-color and attend-to-orientation were performed.The result of six-category classification performance using linear discriminant analysis showed that the accuracy under attended condition was significantly higher than chance level,and significantly higher than that of unattended condition.The accuracy of stacked auto-encoder network was significantly higher than chance level not only for attended condition,but also for unattended condition,but there is no main effect between these two coditions.The SNR feature showed better performance in six-category classification than that of power spectrum feature,but behaved worse in binary classification in this experiment.By designing dual visual stimuli experimental paradigm,the paper performed multivariable pattern analysis for EEG signal analysis,and proposed two kinds of method to extract signal features of corresponding visual characteristic for the experiment.This can provide a reference for muitl-variable pattern analysis of complex visual stimuli...