In order to improve the blurred image registration properties of image registration algorithm, an adaptive blurred image registration algorithm fused nonlinear scale space and space cosine similarity is proposed. The algorithm adopts the nonlinear scale space theory to extract the local feature of the image, and the feature points of the image are extracted by the KAZE algorithm to form the MSURF feature descriptor. Using the space cosine to match the feature points, and match the adaptive threshold according to the different characteristics of the image to seek the optimal transformation relations of reasonable quantity of matching points. Finally, the RANSAC algorithm is used for excluding mistake matching points to improve the accuracy. The experimental results show that the algorithm can effectively improve the matching accuracy and precision of the fuzzy image registration, the matching accuracy and precision can be improved up to 25% and 7.909 pixels, and has better registration performance.