A RANSAC-based cylindrical image registration algorithm was presented. First, the algorithm makes a coarse matching for the detected Harris corner points by using the NCC algorithm. Then, it reduces the false matched pairs by the two pass of improved RANSAC algorithm to increase the number of right matched pairs. At last, it optimizes the affine transformation model parameters by using Levenberg-Marquardt algorithm to lower the registration error. Experiments have shown that the right matched ratio of Harris corner reaches 99.2% after one time of improved RANSAC, and reaches 99.6% after two times of improved RANSAC. Compared with the existing methods, it has achieved higher right matched ratio under similar conditions.