Somewhere around 1997 (I think), Robert Silvers came out with a coffee table book of what he called Photomosaics - a big picture of something made out of tiles of a whole mess of smaller somethings' pictures. I was completely taken aback by it. He also has a website at photomosaic.com. The only problem with his ingenious (and relatively stunning) form of art is that he refused to release the code for the program he invented to make such works.

After a semi-grueling computational physics course, I learned a bit about Neural Networks, which, among other things, are used heavily in pattern-recognition programs. I decided to put this to good use and see if I can't come up with some sort of rival to the original photomosaics, despite my ineptitude at general programming (my parts of the programming work will primarily be in Mathematica, and mostly pure mathematical operations on sets of data). I have a general understanding of how to essentially take a pattern of numbers, and compare it to a list of other patterns, choosing the one that's the closest. With this knowledge, the next leap of intuition is to take the fact that pictures, on a computer, are essentially long lists of 1's and 0's. I can use these lists for my patterns.

The catch here is that there are a lot of things that will need to be done to make this thing work.

I'm going to need to:

• input the patterns of 1's and 0's from a bunch of pictures to make a database to work with
• maybe find a way to store the dimensions of the rows and columns
• flatten down the rows and columns to one long, nasty list
• store the patterns into an ungodly matrix of weighted averages of values corresponding to all the pictures
• convert the list of 1's and 0's into a list of 1's and -1's
• run a disgusting loop that lets the pattern evolve to its nearest relative in the "database" matrix - this is where the most complicated work is done
• take the evolved list out from that loop
• change it back to 1's and 0's
• re-inflate it to have rows and columns
• write it to an output file

To make this work, I'll need assistance on the generalized programming side - my "associate" in this will be doing a chunk of work to make my work work (hehe).

He'll be making things to do most of the grunt work in this, but ridiculously important and strenuous grunt work (at least as far as my non-programming mind knows):

• Convert a picture (bmp file) into a text file chock-ful of numbers
• Convert my output numbers back into a picture

And hopefully, after a little testing period:

• Find a way to break up a big picture into a whole mess of smaller, equally sized pictures that can match up in size to whatever I'm using in the database
• find a way to piece together all of the little pictures I make up back into one big one
• maybe work on the possiblity of color usage
• Hopefully, find a way to notify me of negative images (I can expect half of my outputs to actually be an anti-pattern - not sure what effect that will have on the pictures yet)
• maybe form some sort of macro to run the process over and over for a whole mess of pictures, so that one big picture can be input, and after an unholy-large processing time, a picture can be output...

To accomplish all this, it'll take who knows how long - probably not too much, once we get the ball rolling.

I plan on starting relatively simply - I'll take one small b/w picture, compare it to three equal sized b/w pictures (one of which is very intentionally the same picture with a little defacing - so it should pick that one), and see what happens - if that works, the next step might just be to build up something of an image database...dunno yet

if that part of this "noble experiment" runs scot-free, the rest might just start flying by - check back for updates perhaps...

Latest Update! (ok, not the latest one anymore)

Finally, success on the trial run of the idea

I successfully managed to get the program to choose from three pictures:

and asked it to choose the one closest to my tester:

whereby, it correctly chose the (undefaced) Tick