Among the dimensionality reduction methods of 2D image, they often use L2 norm and L1 norm to construct the dimension reduction model, to a some extent, they realize the goal of dimension reduction. However, these methods only apply to the single norm, which is very limited. In this paper, we propose a new method, which uses Lp norm (1≤p≤2) instead of a single one. This method can construct a more generalized dimension reduction model and is suitable for all similar models. p=1 and p=2 can be regarded as a special case of this model. The proposed method is more flexible than a L1 or L2 norm and can adapt to different problems. In this paper, the objective function adopts the Lp norm of the twodimensional maximum margin criterion, and the intraclass discrete factor is defined to reduce the original data. The experiments on ORL and Yale and noise reduction databases show better robustness and efficiency.