As the Non-negative Matrix Factorization (NMF) algorithm has a higher iteration time complexity and the Gradient Projection(BP) optimization method can significantly reduce the NMF iteration time complexity, an image fusion algorithm by combiing the improved PGNMF and Nonsubsampled Contourlet Transform (NSCT) is proposed in this paper. Firstly, the reistered oriinal images are decomposited in multi-scale and multi-direction decomposition in NSCT domain. According to the characters of the different areas, different fusion rules are designed in the NSCT domain. The low-pass sub band coefficients used as original data impose to the IPGNMF algorithm,to obtain the fused low-pass sub band coefficients and the band-pass directional sub band coefficients impose to the Neighborhood Homogeneous Measurement (NHM) algorithm,to obtain the fusion band-pass directional sub band coefficients. Finally, the fused result is obtained through inverse NSCT. The propose algorithm has been experimented on two groups of different scene images,and experimental results show that it superior to those conventional fusion methods based on NSWT,IPGNMF and NSCT in subjective and objective standards.As contrasted with NSCT method in two groups images, its entropy, definition and QABIF have been increased by 0.0627%、0.901%、3.1201% and 2.769%、2.203%、1.049%,respectively。