elfs | I am geekily giggly dirty minded this morning!

Yesterday,

fallenpegasus turned me on to Ariel Shamir's research, entitled Seam carving: content-aware image resizing, which is a fascinating algorithm for removing "seams" (meandering columns or rows of pixels) from an image in a way that retains the most semantically meaningful portions of the image. The technique is fairly geeky, involving calculating gradient paths to take out the seam that would cause the least loss of information, but it seems straightforward and I have at least two publicly available algorithms (one in Java, one in C) for it along with the paper.

My intent is to put it into my work queue. I used to be a big contributor to the NetPBM suite, although the only thing of mine that remains there is the -c flag in giftopnm, my contribution back from my time working with The National Center for Missing and Exploited Children. (They used to send their "Have you seen me?" cards to the web admin as a scanned GIF with all of the case information attached to the image's comment field.) I'd hoped to put seam carving into the NetPBM suite as well.

While I was researching it, I came across another paper by Jamie Hayes and Alexi Efros entitled "Scene Completion Using Millions of Photographs." The idea here is a simple one: if you cut out or damage part of a photograph, you can "restore" some semantic meaning to the damaged part by culling through millions of other photographs looking for something that would seamlessly fit into the damaged portion. Hayes and Efros have created an algorithm that apparently does this very well. The example at the top of the page is quite remarkable.

It's probably revealing that my first thought while reading the abstract was, "Oh, cool. It shouldn't be too hard to create a detect pixilated region algorithm. If I use that to 'damage' the picture and the Scene Completion algorithm to repair it, I can restore an awful lot of important, uh, semantic detail to all those censored images on the a.b.p.e.asians.* newsgroups."

Hell, 99.44% of those images are photoshopped anyway-- not counting the pixellation. Nobody's going to care if her pubic hair is borrowed. The Internet views censorship as damage and routes around it; all I'm doing is trying to help.

It occurs to me that this sort of digital merkin would have two other significant uses: done right, it might scramble watermarks or steganographs, and it might also remove advertising bugs^[?].