Of course it doesn't really provide any new information, it just creates a realistic result. Which is probably fine for restoring old movies, but not for hunting down murderers.
Anyone who has ever worked with image files knows that, unlike the fictional world of shows like CSI, there’s no easy way to take a low-resolution image and magically transform it into a high-resolution picture using some fancy “enhance” tool. Fortunately, some brilliant computer scientists at the Max Planck Institute for Intelligent Systems in Germany are working on the problem — and they’ve come up with a pretty nifty algorithm to address it.
“The task of super-resolution has been studied for decades,” Mehdi M.S. Sajjadi, one of the researchers on the project, told Digital Trends. “Before this work, even the state of the art has been producing very blurry images, especially at textured regions. The reason for this is that they asked their neural networks the impossible — to reconstruct the original image with pixel-perfect accuracy. Since this is impossible, the neural networks produce blurry results. We take a different approach [by instead asking] the neural network to produce realistic textures. To do this, the neural network takes a look at the whole image, detects regions, and uses this semantic information to produce realistic textures and sharper images.”
To train their algorithm, the researchers fed their neural network a large data set of images to build up its knowledge of different textures. The neural network only gets to see downsampled versions of the images, and is given the task of upsampling these pictures. Once the network produces an output image, the researchers then compare it with the original high-resolution image and tweak the algorithm to correct any errors, such as making it produce sharper edges or more realistic grass textures where it has not done so. After a while, the algorithm is able to do this on its own with no human intervention necessary.
https://www.digitaltrends.com/cool-tech/algorithm-low-res-high-res/
...or looking around a corner into the bathroom mirror to see a tattoo...
ReplyDeleteThis reminds me of the actors who would get turned into police repeatedly after doing a recreation of a crime on "Unsolved Mysteries" or "America's Most Wanted"
ReplyDelete"Sorry sir, your face matches the clip-art our algorithm used to reconstruct detail in grainy surveillance footage. You're sure you have an alibi for November 22nd, 1963?"
As I wrote elsewhere: the kicker is in the second-to-last paragraph: the training set is the original photos with similar content. And I'd bet serious money the training set isn't large. How would it do if trained on the totality of photos uploaded to FB/Instagram/Snapshat/G+/Flickr/etc. over a period of one year? For any reasonable model size (which means throwing away the vast majority of learned features), I'd guess the results on arbitrary photos thereafter would be fairly bad.
ReplyDeleteThere have been other recent ML papers along these lines (e.g., one that trained exclusively on a few hundred celebrity faces, with the obvious result that the artificially generated ones were preternaturally good-looking, and another just this week that trained on a large corpus labeled images like "person," "bridge," "horse," "dog," "car" and created some truly hideous mashups). You can do interesting things, and in some specific cases you can make great improvements over the previous state of the art, but there's no magic here--what you put in is very strongly tied to what you get out.
I suspect the next stage beyond these simplistic approaches will involve modeling the low-res scene in 2D or 3D, identifying and transforming the different elements into some canonical form, applying the ML algorithm to that, and then doing the reverse transformation to produce the final result. IOW, much more emphasis on the "I" in AI.