One of the main problem in digital image processing is image registration, which consists in finding the transformation, which maps points of one image into the same points of another. For computational cost of image registration decreasing parameters of interframe distortions estimate not in every point of image, but only in local fragments located around specified nodes of some regular or irregular grid. Registration parameters between nodes are predicted using estimated registration parameters of local fragments. As these parameters estimated nodes locations (translations) on data image are commonly used, which gives a low accuracy of the registration parameters prediction between the nodes. New prediction methods of parameters between two nodes of local fragments centers grid in case of unknown global geometric spatial distortion model are discussed. These methods let us increase the accuracy of image registration. They are based on using larger number, in comparison with traditional, of estimated parameters of the fragments location. Using of various sets of translation, rotation angle and scale parameters. It was found that the most preferred methods of prediction is based on Bezier and rational Bezier curves. However, their use is complicated by the need to find the parameterization coefficient. Introduction of additional parameters as which scale estimates are applied, provides a little more exact result for a rational Bezier curve. Using of a biarc curve gives quite good results of a prediction also, but here it should be noted that there is the computation complexity increasing which connected with transformation to local coordinates system. Slightly less accurate results, but at a much lesser extent of computational cost is obtained by using a cubic polynomial for prediction. The given experiments showed that, for example, compared to traditional cubic interpolation using four points, the prediction accuracy increases by about an order of magnitude.