For the weakness of maximum margin criterion and margin fisher analysis in the process of human face feature extraction, this paper presents margin discriminant projection algorithm. We define within-class scatter matrix using class samples’ mean and it’s marginal samples of same class, and define the between-class scatter matrix using class samples’ mean and it’s marginal samples of other classes. At the same time, the maximum margin criterionis used to solve singularity of within-class scatter matrix. Compared with the classical maximum margin criterion and margin fisher analysis algorithm, margin discriminant projection can consider the globaland local structure of samples at the same time, avoid small sample problem. The experiments on the face datasets show that the margin discriminant projection is a kind of effective feature extraction algorithm and has improved face recognition accuracy.