Improving the performance of the sparse representation Based classification in BCI systems, by enhancing the process of feature extraction and using an optimized sparse solution finding algorithm

In recent years, Brain-Computer Interface (BCI) has been noted as a new means of communication between the human brain and his surroundings. In order to set up such a system, the collaboration of several blocks, such as data recording, signal processing and user interface are needed. The signal processing block, includes two units of preprocessing and pattern recognition. Pattern recognition block itself involves two phases: feature extraction and classification. In this paper, the sparse representation based classification (SRC) has been used in the classification block. There are two important issues in using the SRC. These are creating an appropriate dictionary matrix and adopting a proper method for finding the sparse solution for an input data. In this research study, the dictionary matrix is formed by extracting an optimal set of features from the training data. Toward this goal, the common spatial patterns algorithm (CSP) is first used. Sensitivity to noise and the over learning phenomena are the main drawbacks of the CSP algorithm. In order to remove these problems, the regularized common spatial patterns algorithm (RCSP) is employed. In previous studies in within the BCI framework, the standard BP algorithm has been used to find a sparse solution. The main disadvantage of the BP algorithm is that the method is computationally expensive. To overcome this weakness, a recently proposed algorithm namely the SL0 approach is used instead. Our experimental results show that when the number of training samples is limited, the RCSP algorithm outperforms the CSP one. Using the features derived from the RCSP, the average detection rate is in average increased by a factor of 7.53 %. Our classification results also show that using the SL0 algorithm, the classification process is highly speeded up as compared to the BP algorithm while an almost equivalent accuracy is achieved.