If you can get games in machine readable format, it may be possible to develop a ML approach that figures out the most likely identifications based on relative positions and as much information as image processing provides.