Sadly I don't think there is too much commercial value in extraordinarily CPU-intensive lossy image compression but it sure beats the heck out of 99.5% of what I saw in undergraduate AI/algorithms-type courses as a demo.
Incidentally, I think from a marketing perspective that is a great test case to use for it. If you had picked, I don't know, some portrait by an unknown artist it would have been no less technically impressive but it wouldn't spread NEARLY as far. (This is one reason why I managed to work the KJV Bible and the US Constitution into my test corpora for almost every natural language processing project I've ever done.)
Most of the papers one sees about serious, potentially lucrative applications of biologically inspired optimization metaheuristics (algorithms like the one in the OP) are engineering-oriented.
For example, NASA used evolutionary computation to design a little antenna thing, and it went into space, and it was human-competitive (as good or better than anything a human could design in a reasonable amount of time). http://ti.arc.nasa.gov/projects/esg/research/antenna.htm
Most of the applications, in other words, have little to do with the kind of highly data-driven web applications with rapidly changing requirements that seems to occupy a lot of the time of the users on this site. :( Still, excellent candidate for spare-time/weekend hacking.
Finally, one basic thing to observe about most of these heuristics is that they are embarrassingly parallel, which means they really could make good use of a 16-core machine.
"Sadly I don't think there is too much commercial value in extraordinarily CPU-intensive lossy image compression"
First, as commenters on that post pointed out, this is horribly inefficient. It's not even a GA, it's a blind hill climber! Hint: you need crossover; calling something "DNA" is not sufficient.
Second, even if it's a large image, and takes say 10 minutes on a laptop-class computer, that's 1 cent on EC2. Computing power has gotten ridiculously cheap.
Third, here is at least one killer app: mobile. B/W still sucks, and rendering speed could use some help as well. I can certainly imagine large websites benefitting from this. This is basically a raster to vector converter. You could start using this right now in SVG enabled environments.
Finally, I just thought of something truly sick. Note that you can draw arbitrary polygons using javascript. http://www.uselesspickles.com/triangles/demo.html Put these two together, and... dear God, I don't even want to think about it.
Plus, the algorithm could always use some optimization. I figure implementing a true genetic algorithm (several 'painters' working at once, recombining their DNA instead of overwriting) would get to a 'good enough' result more quickly.
It makes me wonder whether there is some demoscener who has already done something like this to fit the Mona Lisa (or probably something geekier, like a picture of Clive Sinclair) into their 4k demo.
Related techniques could be utilized for creating non-photorealistic/stylized renderings of an existing/reference image. There has been some work on this.
Sadly I don't think there is too much commercial value in extraordinarily CPU-intensive lossy image compression but it sure beats the heck out of 99.5% of what I saw in undergraduate AI/algorithms-type courses as a demo.
Incidentally, I think from a marketing perspective that is a great test case to use for it. If you had picked, I don't know, some portrait by an unknown artist it would have been no less technically impressive but it wouldn't spread NEARLY as far. (This is one reason why I managed to work the KJV Bible and the US Constitution into my test corpora for almost every natural language processing project I've ever done.)