Image registration is to find the correspondences between different images of the same scene. It still remains a challenging problem especially when there is dramatic changes of object appearances. This paper presents a new image registration method for alleviating this problem, which first finds sparse correspondences and interpolates continuous dense motions from them. Unlike conventional registration methods, we regularly place control points to prevent biased distribution of feature points and find optimized correspondences by minimizing the cost function. The cost function is based on SIFT descriptors and considers the smoothness of motion and topological relations. Especially, the topological term prevents inconsistent solution like fold-over or duplication artifacts. For the optimization, we adopt a dual-layer belief propagation and coarse-to-fine scheme. Based on barycentric coordinates, we finally estimate dense motion from sparse correspondences. Experimental results show that the proposed method yields more plausible results and is computationally efficient.