The registration of retinal images is required to facilitate the study of t
he optic nerve head and the retina. The method we propose combines the use
of mutual information as the similarity measure and simulated annealing as
the search technique. It is robust toward large transformations between the
images and significant changes in light intensity. By using a pyramid samp
ling approach combined with simulated reannealing we find that registration
can be achieved to predetermined precision, subject to choice of interpola
tion and the constraint of time. The algorithm was tested on 49 pairs of st
ereo images. and 48 pairs of temporal images with success.