Single-trial ERPs Denoising And Classification Based On Sparse Representation

Event related potentials (ERPs) is the evoked potentials when human visual, auditory, or feeling is suffering from stimulation. As an important means of researching the human brain cognitive function, the ERPs has been widely used in cognitive neuroscience and the study of clinical medicine. In order to detect and treat brain diseases, it is important to efficient extract and accurate judge the ERPs signal from electroencephalogram (EEG). However, the signal to noise ratio(SNR) of ERPs is very small since the ERPs signal is very weak and often superposing with the background noise of EEG signal. Conventional methods, which are based on the assumption that the background noise is stationary random processes, often obtain the ERPs signal by adopting superposition and average (SA) of single experiment signal through many times. However, SA-based approaches not only can’t meet the precondition in practice, but also cause the important cognitive information loss in every single experiment.The basic idea of signal sparse representation, which has been widely applied in various fields of signal analysis and processing, is to reconstruct the original signal by making use of the non-zero coefficient as little as possible. In this paper, some ERPs signal denoising and classification algorithms, which can overcomes the shortcoming of existing classical algorithms, is proposed on the basis of sparse representation theory. The main research contents and results are as follows:1. Traditional SA-based approach is not reasonable in dealing with the aspect of noise assumption and cause the cognitive information loss in every single experiment. In order to overcome the aforementioned drawbacks, two novel algorithms based on Non-local Means and collaborative filtering are proposed in this paper. Firstly, the ERPs image is combined by single experiment signals. Secondly, the block operation is adopted to utilize the similarity between these blocks, which is used to reconstruct the signal in the framework of sparse representation. Experimental results demonstrate that the proposed two algorithms can overcome the shortcomings of conventional algorithms, meanwhile the denoising effect is satisfactory. Moreover, the computational time is also improved while comparing with classical algorithms.2. The dimension of ERPs data is usually ten thousand. This makes the classifier, which based on the discriminative function, can only judge and analyze the ERPs signal by extracting the its features, and the feature extraction is directly related to the accuracy of the classification results. It is difficult to select the discriminant features from the ERPs signal for its large data dimension and hard to meet the number requirement of variable of classifiers. Meanwhile the low signal to noise ratio also affects the accuracy of feature extraction. In this paper, the proposed sparse representation based classification (SRC) algorithm, the discriminant K-SVD algorithm and the label consistent K-SVD algorithm focus on the minimum reconstruction error in the sparse representation framework instead of the feature extraction. Experimental results show that the proposed algorithm has a better performance on classification effect and computational time while comparing with conventional algorithms. Moreover, the problem of sample shortage in ERPs signal acquisition processing can be solved by the proposed method.