Registration based on mutual information is a typical method in medical image registration. Mutual information is a common similarity measure in image registration, which has excellent robustness and accuracy, but large calculation amount makes it difficult to be applied to clinics. A maximization of mutual information based image registration method is described. Firstly Because of using maximum mutual information to image registration have inferiority, the registration based on local curvature maximum to obtain corner points. Then, the one to one matching points could be obtained through mutual information rough match. Experimental results indicate the proposed algorithm can achieve better accuracy and good robust.