What an interesting idea. I completely see your point. As a computer programmer, I see how easy it is to make the code; and that it needs SERIOUS
refinement before it's "usable"...so I start thinking "how could it be improved"?
Here are my answers to that last:
* A number of different "refinements" will be needed to provide usable images...maybe as early as 4 or 5 generations from now!
* start with black-and-white images
* start with a small image space (the 2x2 image makes sense)
* multi-tasking will probably help a lot...
* have computer-aided recognition of "candidates"
+ facial recognition
+ geographic recognition (don't know if such a thing exists at the moment)
+ spacial (yes, spacial...as in pictures of space...I'm guessing it's a made-up word
+ scale down a library of "known good" pictures (say, every scene for a movie, or a video game...or a collection of such; maybe even include
artwork?) and use these as "exemplars" and set a "similarity threshold" (this should not be used on it's own, because it will probably strip out
"pictures" that are legitimate, but unlike anything imagined or captured by a camera)
* when "candidates" are found:
+ show them to a human being and use that feedback with a neural net to refine the definition of "useful"
+ run them through a "refiner" - add colors, pixels or both
+ run "refined" images through the "candidate selection" and the "when found" mechanisms.
With today's computers, I'm pretty sure we could actually process 10 images per second (that's a conservative estimate...I'm sure we're not far
from being able to process 100 images per second!)
Just adding in my thoughts for a "better way" to do this. Strictly a thought experiment for me...I'm too lazy and too busy with other ideas to do
anything about it!