So, I'm interested in learning how to detect patterns in a set of images and then use those patterns to create a new image of similar style. For example, say there is a group of 20~ish images (hopefully a large enough sample) with some interesting characteristics and one wished to produce a new image of a set size with such characteristics for an application's background. The catch is that the background should be different each time it is loaded, so it must be generated - not produced in an image editing software.

Now, once I've the base concepts down, I should be able to write something genetic to spit out images until I like the results, but I'm not sure how to start researching this one. I can't really use fractal algorithms because I want to avoid the repeating behavior. So, which metrics do you think I should obtain, and do you know of any good resources for using said metrics to create new images that incorporate the patterns/tendencies of an image set?

  • $\begingroup$ It is not a difficult task but needs quite much effort to get such a good algorithm. There are quite many segmentation algorithm with which you can determine similar contents. What do you mean by metrics? can you give an example or elaborate on this? $\endgroup$ – Seyhmus Güngören Aug 14 '12 at 14:41
  • $\begingroup$ "Similar style" is a very fuzzy concept to nail down, but you might find it interesting to follow some links from here. $\endgroup$ – Henning Makholm Aug 14 '12 at 14:44
  • $\begingroup$ There are several image metrics (none of which I am too familiar with), such as line profiles, hue/saturation measurements and greyscale intensity. I'm sorry about the vagueness of 'style;' I'm asuming that after deciding what I need to measure to get pointed in the right direction, I'll need to plug the values into a genetic algorithm to 'nudge' the results into the exact output I'm looking for. Thanks for the links; this will make a great start! $\endgroup$ – bendicott Aug 14 '12 at 17:13

You can start with reading and implementing these two papers: "Texture Synthesis by Non-parametric Sampling" Alexei A. Efros and Thomas K. Leung IEEE International Conference on Computer Vision (ICCV'99), Corfu, Greece, September 1999, and J Portilla and E P Simoncelli. A Parametric Texture Model based on Joint Statistics of Complex Wavelet Coefficients. Int'l Journal of Computer Vision. 40(1):49-71, October, 2000.

  • $\begingroup$ Awesome, these look great. Thanks for the help! $\endgroup$ – bendicott Aug 14 '12 at 17:17

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.