Hi Wayne,
Before computing the cross-correlation, both images need to be zero-padded to the same size. This ensures that the frequency-domain multiplication is valid and that the result captures the full cross-correlation.
Once the images are appropriately padded, you can compute the cross-correlation using:
c = ifft2(fft2(X) .* conj(fft2(Y)));
where X and Y are the zero-padded versions of the original images.
I have used the approach described in this article for one of my own projects and found it very effective for template matching:
I think you'll find it helpful as well.
