Found this today last month. If you click on the picture, you’ll see that there are parts that are composed of other images.
I prepared a folder with full of small pictures, and a preprocessor scanned all files in that folder, took their R G and B averages and saved it to a database.
I then divided the big picture into small chunks (the small, the better, but also requires more photos in order not to have repetitions). I took R G B averages of each part and computed the distance from each of the items stored in DB(for efficient access, they are fetched to memory). The distance formula is simply the Pythagoras theorem.
![clip_image002[6]](http://tunatoksoz.com/image.axd?picture=clip_image002%5B6%5D_thumb.gif "clip_image002[6]")
The closer it is, the better it fits to the place. However, there is one problem. Neighbor parts tend to have a correlation (they share similar colors), and if we use the same picture for neighboring cells, it doesn’t look good(above probably left from an early version). What I did was to take closest 10 and select randomly.
The rest is a bit of Image processing(not processing but replacing, lets say)
I had fun doing it :) Art is everywhere even if I don’t think I like it :)