Can such a model be used to compare a real picture of a suspect with a low res picture and produce an objective measure of similarity (i.e. not only raw pixel s

Can such a model be used to compare a real picture of a suspect with a low res picture and produce an objective measure of similarity (i.e. not only raw pixel similarity but a similarity measure that actually takes into account how actual human faces downscale, possibly resilient to minor rotations?)