Nonnegative Matrix Factorization Based on Projected Gradient and Underapproximation Method

In order to improve the ability of the parts-based representations of the Nonnegative Matrix Factorization(NMF) algorithm,this paper proposes a NMF based on projected gradient and underapproximation method——Projected Gradient Nonnegative Matrix Underapproximation(PGNMU).By adding the upper bound constraint,it applies the important features of the basis matrixs that are extracted by using the alternating iterative method based on the projected gradient methods to the experiments.Compared with previously published methods on the CBCL and ORL database,results show that the method has the better sparseness and better recognition rate than the others.