Presents a new non-rigid image registration algorithm based on nonlinear diffusion. Firstly derivates the mathematic model of nonlinear registration with variational method. Then incorporates the local information of image into the mathematical framework and self-adjusts the diffusion coefficient according to image content. So the deformation in different regions obtains varying degrees of smoothness. In order to improve registration rate, adopts additive operator splitting method to solve the partial difference equation. Experiments with synthetic images and MRI images show that the new method achieves better performance than the non-adaptive regularization methods, preserves local details and has better registration accuracy.