It's an extremely hard problem. Much harder than you'd think. Just spotting a known "thing" in an image is often a research problem; doing any generic image is much harder.