An original scheme to automatically generate control point pairs for image registration in a progressive way is presented here. It consists of three stages of processing. In the first stage, the Voronoi-Delaunay dual graph is used to predict the optimal location for a new control point to be generated. Second, a target-defined interest operator is designed to extract feature points from the predicted area. In the third stage, a hierarchical templet matching procedure is applied. A global consistency check by robust estimation to improve its reliability is also included. The major features of the proposed scheme are that (1) it yields highly reliable and accurate results, (2) the distribution of generated control points is uniform, and (3) it is computationally efficient. Examples illustrating the operation of the scheme are presented.