Automatic image registration algorithms that rely on a gradient descent based approach may fail when the initial misalignment between objects is large. The registration task is even more difficult for multi-modal images because of the non-linear relationship between the pixel intensities in the images to be aligned. In this paper we will present a multi-modal image registration algorithm which successfully registers 3D CT to 2D fluoroscopy data for large initial displacements between the images. The approach uses the conditional means (CM) of the joint probability distribution of the images to establish a model linear relationship between the pixel intensities of the images and then applies log-polar transforms (LPT) in the frequency domain to estimate the in-plane scale and rotation changes between the images. Our experimental results show that the proposed approach can increase the range of initial displacements for which the algorithm is able successfully register images by a factor of 4 when compared to the best of the existing gradient-descent based approaches.