In order to overcome the disadvantages of the low registration accuracy and the computational complexity of traditional algorithms, a new sub-pixel registration algorithm based on combination of curved surface fitting method and gradient method was proposed. Firstly, curved surface fitting method based on 9 correlation coefficients was used to achieve the image rough registration. Secondly, subimages were extracted from the reference images according to a predefined selection method. At last, gradient method based on the rough registration was proposed to achieve the accurate sub-pixel registration. Test results among experimental images with different translational relations show that the new algorithm has the advantages of both curved surface fitting method and gradient method, and greatly improves the accuracy of image subpixel registration．The maximum absolute error of registration accuracy is reduced from 0.17 pixel to 0.02 pixel.