NMF has been extensively applied on various pattern recognition problems, including face recognition. To enhance the performance of NMF for face recognition, new additive iterative step sizes are proposed for the basic NMF method, which can raise the searching accuracy during the iteration, and then the recognition rate can be improved. The improved NMF method is named INMF. Meanwhile, the experiments results show that the proposed improved additive iteration can also raise the recognition rate of the SNMF and WNMF. Besides, we find that no sparse constraint is applied to INMF and lots of redundant information still exists, thus a threshold-sparse constraint is introduced to make the base matrix W to a 0-1 matrix, and then the feature data of the base matrix become sparse, therefore the recognition rate can be further improved. The INMF model with the threshold-sparse constraint is named SINMF. Finally, our extensive experimental results showed that the highest recognition rate of the SINMF method can achieve 99%, with improvement over the INMF, IWNMF, ISNMF and deep NMF methods by 11%, 5.5%, 11% and 8%, respectively. Meanwhile, compared with the deep convolutional neural network, the recognition rate of SINMF method is proved more efficient.