I'm back at working with 3D images that have to be registered with an affine registration. I compute the ssd between the moved and the target image as the objective function. If I use linear interpolation to move the images I often risk local minima when using fminunc. If I use cubic interoplation each objective function evaluation takes forever and the minimization becomes not usable in practice.
Does any of you have any suggestion about a practical workaround for this? (i.e.: being able to keep using linear interpolation while avoiding local minima in fminunc? My images are already very smoothed compared to the pixel size) Or any other kind of suggestions?