So I think some sort of hybrid between object recognition (like being discussed here as part of the workflow) and standard image processing stuff could be helpful there. E.g. it's not absolute sharpness that you're looking for it's the subject being sharp (and possibly sharper than in other photos from the same time period of the same subject).