I have the following situation:

I have an image file with each pixel having a specific value. Now I want to rearrange and distribute these pixel values so that the result image looks similar to a set of other images that share a similar pixel value distribution.

I wanted to know how I should tackle this challenge. I know basics in machine learning and would "just" use the training images as a csv file and create a model from it to apply to my other image. Any tips you could provide me?

  • $\begingroup$ Take a look at this: codegolf.stackexchange.com/questions/33172/… $\endgroup$ Sep 14, 2019 at 8:31
  • $\begingroup$ @NeilSlater thanks for the link. Looks really interesting. I will take a deeper look into ot later, but it seems that it may not be exactly what I want to achieve. There, the goal was to redistribute pixels to get a specific other image, so a algorithm "just" looked for similarities in the target and assign the pixel from the source to this. However, I don't want to imitate a specific target image but rather want to get an image that looks similar (e.g. from a set of face images I want to create another image of a face that is not a duplicate of another one given a random set of pixels). $\endgroup$
    – JoeBe
    Sep 14, 2019 at 13:55
  • $\begingroup$ I think you need to define "similar" in order to get help with that part. I took "similar" to mean exactly what the link does (because clearly it is not identical if it takes colours from another image). Be specific. What about the set of other images counts as being similar for your purpose? E.g. showing some common set of foreground objects? Having same number of identifiable objects? You cannot just go for "visually similar" since that is poorly defined. $\endgroup$ Sep 14, 2019 at 17:38
  • $\begingroup$ Ok. Let's assume an example where I have only black and white pixels and I want to create a labyrinth from it. I have a set of hundreds of complex labyrinths, where white stands for a path and black for a wall. Let's also assume that a player who loves labyrinths already knows all of these labyrinths. So he wants something new. I also have a set of a random number of black and white pixels. Now an algorithm / model should rearrange these pixels to create another labyrinth (thus similar to the others) but with a new set of paths $\endgroup$
    – JoeBe
    Sep 14, 2019 at 18:37
  • $\begingroup$ Generating random labyrinths has lots of solutoons that don't involve manipulating pixels, and generating from images only would be a really hard problem, you would need millions, not hundreds, of examples, for a ML-based solutoon to work (ML solution won't "get" key features of labyrinths such as need for a clear path). I suggest you ask about what you really want to do, as opposed to try and helpfully simplify it to get well-defined answers, because you may get answers you cannot adapt to your use case. But you do need to be specific about terms such as "similar". $\endgroup$ Sep 14, 2019 at 19:25


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.