The Extraction And Classification Algorithm Of Visual Evoked Potentials

Visual evoked potentials can reflect the lesions of the human visual pathway. Clinically, it’s an efficient and simple mean for detecting the visual pathway diseases. The conventional method of extracting potential is superposition average, in order to obtain a desired potential, the results usually be superposed by thousands of times. The repeated stimulation in patients may prone to fatigue of nervous system and the evoked potential is difficult to obtain the desired signal quality. So in recent years, how to reduce the stimulation frequency, shorten the measurement time become the main content of the study. In addition, make it more convenient and effective to realize the real-time tracking and classification of evoked potentials also become a hot topic.In terms of signal filtering, from the sub-space method, principal component analysis to modern wavelet analysis, independent component analysis algorithm, adaptive filters, etc. Have been used in the extraction of evoked potentials. Through experiments and simulation of researchers, based on the advantages and disadvantages of each method, a comprehensive of two or more methods to get potential achieved a good result. In terms of signal classification, neural network classification, decision tree classifier, linear discriminant analysis, support vector machines classifier, K-nearest neighbor classifier and other methods have been well used in biological signal recognition for brain signals.The paper on the basis of previous studies, based on the traditional adaptive noise canceller and the adaptive filtering principle, use of rapid convergence recursive least squares (RLS) algorithm to isolated VEP signal from background EEG; then using wavelet thresholding method to determine the optimal decomposition level. Making denoising results achieved better visual effect by an improved threshold function.To achieve further classification of normal and abnormal VEP signals, choose cataract patients’ VEP signal as abnormal signal source. According to the most representative VEP waveform characteristics, using wavelet multiscale decomposition, select the average of wavelet coefficients and energy of each scale as the characteristic parameter, send it to support vector machine to train and classify. Analysis the classification accuracy rate that affected by different number of sample sets and kernel functions, and using cross-validation method to get parameter optimization to improve the classification accuracy.