The alignment of an image pair by means of transformation matrices is a challenging problem in computer vision, especially when the transformation is unavailable to compute and free of noise. In this paper, we first use the SIFT algorithm to detect the feature points, and then during the image alignment process, we propose to iterate the reference images and solve the transformation matrix under the least squares constraint. Experimental results on aligning some real images with few or none overlap regions show the proposed alignment algorithm outperforms the stateoftheart image alignment methods.