hmmm you could try throwing “saliency” algorithms at those pictures. they basically highlight where interesting/non-background/contrasting stuff is in a picture. that should take care of deciding where a coin is and then also if it’s lighter or darker than the rest of the picture (which way to threshold).
since all of those are quite circular, perhaps also throw a Hough transform for circles at these pictures. it’s gonna have trouble seeing ellipses but if you need that there are extensions and related methods (fast radial symmetry transform, etc)