A new method based on the sharpness-constrained non-negative matrix factorization (SNMF) technique was presented for multi-focus images fusion. An new objective function was defined to impose sharpness constraint, in addition to the non-negativity constraint in the standard NMF. An algorithm was presented for SNMF. The local feature based representation could be obtained by choosing suitable dimension of the feature subspace in NMF. It was pointed out that when using SNMF, if the dimension of the feature subspace was set to 1, the resulted feature base was just the fusion result of the original input images. The feature base obtained included the global feature of the original images. Experimental results were presented to compare SNMF with wavelet transform and Laplacian methods for image fusion, which demonstrates advantages of SNMF in preserving the global feature information.