Abstract: In this paper, we use the "back" sampling window to divide the neighborhood into four sub-domains and use the adaptive quantization strategy to divide the gradient histogram and the local area. Finally, we construct 96-dimensional feature point descriptor, and further improve the SIFT matching algorithm Robustness and real-time. Based on the improved SIFT algorithm, this paper proposes a new image fusion method based on Contourlet transform theory, which solves the fusion problem of edge information and texture information, and eliminates chromatic aberration to make the fused image clearer. Experimental results show that this new method can speed up the SIFT process in image registration and enhance the image stitching effect.