Aiming at the low recognition accuracy of non-negative matrix factorization (NMF) in practical application, an improved spare graph NMF (New-SGNMF) is proposed in this paper. New-SGNMF makes full use of the inherent geometric structure of image data to optimize the basis matrix in two steps. A threshold value s was first set to judge the threshold value of the decomposed base matrix to filter the redundant information in the data. Using L2 norm, sparse constraints were then implemented on the basis matrix, and integrated into the objective function to obtain the objective function of New-SGNMF. In addition, the derivation process of the algorithm and the convergence analysis of the algorithm were given. The experimental results on COIL20, PIE-pose09 and YaleB database show that compared with K-means, PCA, NMF and other algorithms, the proposed algorithm has higher accuracy and normalized mutual information.