"Got an image enhancer that can bitmap?"

July 13, 2012 1:28 PM

Super-Resolution From a Single Image presents interactive examples from a 2009 study of methods for increasing the resolution of digital images.

For maximum enjoyment, click the blue "Our SR Result" buttons beside each image while saying "Enhance!" in a commanding tone.

Via Hacker News.

posted by oulipian (19 comments total)

Web site is really slow right now. I couldn't download the .pdf, and many of the enhanced images failed to load, leaving me clicking the button and trying to see the (non-existent) difference. I'll go back later; I'd really like to know how they do it. (My first guess was "propagating extra edge information.)
posted by benito.strauss at 1:48 PM on July 13, 2012

I didn't really appreciate this until I got to the eye chart... Wow.
posted by polywomp at 1:50 PM on July 13, 2012 [1 favorite]

I waited around for the site to load and got it to work. It's neat. But, I'm also seeing a ton of banding/posterization going on. The portrait of the woman with all the curly hair shows this off especially well, as does the first picture of the baby.
posted by Thorzdad at 1:51 PM on July 13, 2012 [2 favorites]

Enhance!
posted by mrnutty at 1:59 PM on July 13, 2012 [1 favorite]

saying "Enhance! " in a commanding tone.

Oh sonofa...
posted by mrnutty at 2:00 PM on July 13, 2012

Yeah, I got to the picture of the curly-haired woman and thought "that's going to give them problems".
posted by benito.strauss at 2:06 PM on July 13, 2012

That's pretty impressive. There is some sort of posterization going on there, but given the size of the images they are enhancing, it's not that bad.
posted by PipRuss at 2:13 PM on July 13, 2012

Coral Cache of the first link that may work better for some. This is cool!
posted by xedrik at 2:16 PM on July 13, 2012 [1 favorite]

I'll go back later; I'd really like to know how they do it. (My first guess was "propagating extra edge information.)

My guess is reticulating splines.
posted by Philosopher Dirtbike at 2:38 PM on July 13, 2012 [3 favorites]

Yeah, the eye chart is the one that best demonstrates what they're doing. Some of the letters are too low-resolution to be recognizable, and this algorithm converts them to not-quite letters (or could be even incorrect letters and numbers). It's looking up the neighborhood around each pixel in a database and replacing it with the best higher-resolution match.

I thought of doing this with satellite images, which would let you zoom in from a low detail map and essentially make up the detail on the fly as you zoom in. Might be cool.
posted by RobotVoodooPower at 2:40 PM on July 13, 2012 [1 favorite]

And to clarify my last comment, in this method they only build their database from a single image -- so in the eye chart example, the detail in the small letters is inferred by looking at the larger letters and comparing them to what they'd look like if they were lower-resolution.
posted by RobotVoodooPower at 2:48 PM on July 13, 2012

It's looking up the neighborhood around each pixel in a database and replacing it with the best higher-resolution match.

I think there's actually something a little more subtle going on, which is that each area is reconstructed using multiple higher-resolution matches, with closer matches being weighted more strongly. As I read it, the authors claim that doing this helps avoid hallucinating really implausible details.

I thought of doing this with satellite images, which would let you zoom in from a low detail map and essentially make up the detail on the fly as you zoom in.

That reminds me of one of the Multiscale Texture Synthesis results.
posted by Serf at 2:53 PM on July 13, 2012 [2 favorites]

Oops, I probably should have linked to the project page instead, since that has videos.
posted by Serf at 2:54 PM on July 13, 2012

I think we killed it.
posted by Scientist at 3:21 PM on July 13, 2012

I rather prefer the Fattel results.
posted by anigbrowl at 3:28 PM on July 13, 2012

It's a cliche, but they should have included an image of a license plate since that seems to come up in so many TV shows. Then again, the thought of this being used in red light cameras gives me the willies.
posted by tommasz at 3:48 PM on July 13, 2012 [1 favorite]

I wonder what would happen if they tried to use one of these enhanced images in a court case. Some of the info that goes into making the enhanced image comes from relatively distant parts of the original image, no?
posted by benito.strauss at 4:24 PM on July 13, 2012

Heh. These are the filters you can put on old games in emulators - bilinear, bicubic, 2xSAI, Super Eagle, etc -- I recognize each one's artifact pattern. Always interesting, and it kind of depends on the image but they seem to have gotten some decent results.
posted by BlackLeotardFront at 5:31 PM on July 13, 2012

These are the filters you can put on old games in emulators - bilinear, bicubic, 2xSAI, Super Eagle, etc

I don't think that's entirely correct. Bilinear and bicubic, yes. But most of the other results shown gather some sort of global image statistics in order to generate their results, while 2xSAI and Super Eagle are (as far as I can tell) exclusively local methods.

It's possible that 2xSAI and its ilk do approximately amount to one of these other methods, given a reasonable guess at the image statistics. (Or a reasonable guess at the statistics of frames from video games of a certain age.) I can't say for sure, though.
posted by Serf at 11:49 PM on July 13, 2012