A locality projection discriminant analysis (LPDA) method by using local structure-based projection techniques is proposed to extract discriminative features in high-dimensional sample space. Two locality metric matrices are defined to make the discriminative projections preserve the intrinsic neighborhood geometry of the within-class samples while enlarging the margins of extra-class samples near to the class boundaries. LPDA efficiently addresses the nonlinear property of data and the small sample size problem in face recognition scenario; moreover, it can reduce the dimensionality of the original data (such as the role of principle component analysis) as well as extract complete discriminative features in dual subspaces. Experiments on synthetic data sets and ORL, PIE, and FERET low-resolution face databases are performed to evaluate LPDA-based methods and some known methods. The results demonstrate the effectiveness of LPDA and reveal some characteristics of this pure local structure-based method.