We present a new use of mutual information as a similarity technique for the registration of images. Unlike prior work we find that no more than two resolutions are necessary to gain sub-pixel accuracy and that classical optimisation techniques such as gradient descent or Powell's minimisation are not appropriate. Our results show that mutual information combined with a low resolution simulated annealing search yields very fast first approximations to the registration over a wide range of rigid transformations. The method is particularly robust in terms of changes in colour, intensity and noise between images. This makes it particularly useful for multi-modal and temporal image registration. The method was applied successfully to stereo and temporal retinal image pairs.