To get a more accurate fusion image with richer information, the Finite Discrete Shearlet Transform (FDST) is used to improve the multi-focus image fusion algorithm. The perfect translation invariance of FDST and its high efficiency of decomposition and reconstruction are used to obtain the coefficient of high- and low- frequency sub-band by multi-scale and multi-direction decomposition. The fusion method of adaptive weighting used together with regional standard deviation matching is used for the high frequency, and the non-negative matrix factorization of improved gradient projection is used for the low frequency, then the high- and low-frequency sub-bands after fusion are obtained respectively. At last, the fusion image is reconstructed by FDST inverse transform. The experiments on multi-focus image show that, with the improved method, the fusion image is more distinct in subjective vision, has better objective indexes with shorter running time. The fusion result not only retains the rich and effective information of the source image, but also has a good effect.