Abstarct: Face is one of most important biometrics for human identification. Human identification using face image is an essential problem in pattern recognition. Despite many attentions received to solve this problem, still there are a number of challenges which cause face recognition being an open area for research. A number of challenges are face recognition when the number of classes increase, when a part of face is covered, aging phenomena and recognition when image resolution or quality of image is low. In recent years, sparse representation has been use for face recognition widely. Learning a proper dictionary for classes is a key issue. Considering that for different persons face patterns are different despite common structure, for improving sparse representation of human faces a specific dictionary for each person is learnt. Moreover, a common dictionary is learnt to model similarities between different persons. One of solutions for improving accuracy of face classification is classifier fusion. In this thesis we divide face image into nine parts. For each part we design a classifier using space representation. The result of nine classifiers then is fused to better determine the class label for the provided face image.