Summary: Aiming at the slow resolution speed of non-negative matrix factorization (NMF) for face recognition and the disadvantage that the space of bases is not suitable for Euclidean distance measuring, a method for face recognition using wavelet transforms and orthonormalized bases of non-negative matrices is put forward. Experimental results show that the presented method achieves a faster training speed and higher accuracy of recognition while having a lower sensitivity to illumination, facial expressions and partial covering.