Recently, Motor Imagery (MI) based on Brain Computer Interface (BCI) systems have been noticed for neuro-rehabilitation methods. The challenging key is to correct detection of MI tasks. Probabilistic Common Spatial Pattern (P-CSP) is the most recent and efficient method for discriminating two classes of electroencephalogram (EEG) from Motor Imagery (MI) Task. P-CSP resolves the overfitting which is the main issue of spatial filters. The accuracy of true detection tasks is related to some initial values, like the number of sources. In this paper, we generate a feature set by extracted features of each unique filter and select discriminant features by sparse dictionary learning method. Optimal sparse regularization parameter is selected by cross-validation on training data and the maximum accuracy with the corresponding parameter is reported for each subject. The performance of the proposed method is evaluated on publicly available BCI Competition IV dataset 2a to detect two pair classes. Our results have been compared with P-CSP, by sweeping the Best number of source and using automatic selection methods. The results show that sparse features selection from feature set outperforms the existing P-CSP in terms of classification accuracy, and reduces the computational time of selecting the best number of the sources by sparse selection.