PixelNN – Example-Based Image Synthesis
cs.cmu.edu
cs.cmu.edu
Now it seems the possibility of that becoming realistic are increasing with a steady clip, based on this paper and other enhancement techniques I've seen posted here.
You fundamentally can't get back information that has been destroyed/or never captured in the first place.
What you can do is fill in the gaps/information with plausible values.
I don't know whether this sounds like I'm splitting hairs, but it's really important that the general public not think we're extracting information in these procedures, we're interpolating or projecting information that is not there.
Very useful for artificially generating skins for each shoe on a shoe rack in a computer game or simulation, potentially disastrous if the general public starts to think it's applicable to security camera footage or admissible as evidence...
And even then, I'm a little suspicious of how close some of the images got to original without being given color information.
It appears that info was either hidden in the original in a way not apparent to humans or was implicit in their data set in some way that would make it fail on photos of people with different skin tones.
> the second stage uses a pixel-wise nearest neighbor method to map the smoothed output to multiple high-quality, high-frequency outputs in a controllable manner.
My interpretation is that they select training data by hand and generate a bunch of outputs. Repeating the process until they like the final result. From the paper:
> we allow a user to have an arbitrarily-fine level of control through on-the-fly editing of the exemplar set (E.g., “resynthesize an image using the eye from this image and the nose from that one”).
I would argue that this is a form of enhancement though, and in some cases will be enough to completely reconstruct the original image. For example, if I give you a scanned PDF, and you know for a fact that it was size 12 black Ariel text on a white background, this can feasibly let you reconstruct the original image perfectly. The 'prior' that has been encoded by the model from the large amount of other images increases the mutual information between grainy image and high-res. The catch is that uncertainty cannot be removed entirely, and you need to know that the target image comes from roughly the same distribution as the training set. But knowing this gives you information that is not encoded in the pixels themselves, so you can't necessarily argue that some enhancement is impossible. For example with celebrity images, if the model is able to figure out who is in the picture, this massively decreases the set of plausible outputs.
http://www.dkriesel.com/en/blog/2013/0802_xerox-workcentres_...
Enhancing some things incorrectly would be worse than leaving them ambiguous.
When humans think about "enhance", they imagine extracting subtle details that were not obvious from the original, which implies that they know very little about what distribution the original image comes from. If they did, they wouldn't have a need for "enhance" 99% of the time -- the remaining 1% is for artistic purposes, which this is indeed suited for.
It'll be interesting to see how society copes with the removal of the "photographs = evidence" prior.
> when enhancing celebrity images, if the model is able to figure out who is in the picture this massively decreases the set of plausible outputs.
This is an excellent insight.
"Zoom! Enhance! Zoom! Enhance! Enhance! Oh my god it's full of Smurfs..."
A good example is with compression algorithms for media. They work because the sound or image is predictable. And they are ineffective when the input becomes more unpredictable. But if the output is all you have then running the decompression will probably be better than just reading the raw compressed data. But you have to be aware of the limitations.
I love this cliché. I've seen it thousands of times, and probably written it myself a few times. We all repeat stuff like that ad nauseam, without ever thinking.
Because it's fundamentally flawed, especially in the context that it has usually been applied to, namely criticising the CSI:XYZ trope of "enhancing images".
The truth is that there is a lot more information in a low-res image than meets the eye.
Even if you can't read the letters on a license plate, it can be recovered by an algorithm. If the Empire State Building is in the background, it's likely to be a US license plate. Maybe only some letters would result in the photo's low-res pattern. If you only see part of a letter, knowing the font may allow you to rule out many letters or numbers etc...
It's similar to that guy who used Photoshop's swirl effect to hide his face, not knowing that the effect is deterministic, and can easily be undone.
The error mostly appears to be in assuming that the information has been destroyed, when in reality it's often just obscured. And Neural Nets are excellent in squeezing all the information out noisy data.
The effect does not only need to be deterministic, but also invertible.
A low-res image has multiple "inverses" (yikes), supposedly each with an associated probability (if you would model it that way). So it would be more honest if the algorithm shows them all.
> I love this cliché. I've seen it thousands of times, and probably written it myself a few times. We all repeat stuff like that ad nauseam, without ever thinking.
It is not a cliche it is an absolute truth. Information not present cannot be retrieved. There may be more information present than is immediately obvious.
> Neural Nets are excellent in squeezing all the information out noisy data
Maybe but they are also good at overfitting onto noisy data (the original article is an example of such overfitting).
Yes, a low-res image has lots of information. You can process that information in many ways. Missing data can't just be magically blinked into existence though.
Copy/pasting bits of guessed data is NOT getting back information that has been destroyed or never captured. Obscured data is very different from non-existent data. Could the software recreate a destroyed painting of mine based on a simple sketch? Of course not, because it would have to invent details it knows nothing about.
I think it's almost dangerous to call this line of thinking cliché. It should be celebrated, not ridiculed.
https://photo.stackexchange.com/questions/17098/csi-image-re...
Somewhat no, but somewhat yes. Thing is, while there can be lots of input images that generate the same output, it could be that only one (or a handful) of them would occur in reality. If this happens to sometimes be the case, and if you could somehow guarantee this was the case in some particular scenario, it could very well make sense to admit it as evidence. Of course, the issue is that figuring this out may not be possible...
But that's not fully accurate either. Sometimes the information in total will really be a more accurate representation of reality than the blurred image. Maybe it could be described as an educated guess, sometimes wrong, sometimes invaluable.
It would be interesting to see the results starting with higher quality images. With the camera quality increasing, many times there should be more data to start with.
A
When is a guess invaluable?
[0] https://i.pinimg.com/originals/b5/29/1b/b5291bba7250abd12010...
Already being done today with DNA.
- I'm joking of course :) hehe
https://www.wired.com/2017/04/courts-using-ai-sentence-crimi...
This thread a year ago worried about it too, but the paper itself seems implausible and problematic.
Don't get me wrong, I think we're still far far far off situation where we can get those reliably, but I can see how you could get the actual face out of a blurred image.
Errr wrong. A perfect hash, yes. But they're never perfect. You have a collision domain and you hope that you don't have enough inputs to trigger a birthday paradox.
Look at the pictures on the article. It's an outline of the shoe. That's your hash. ANY shoe with that general outline resolves to that same hash.
If your input is objects found in the Oxford English Dictionary, you'll have low collisions. An elephant doesn't hash to that outline. But if your inputs is the Kohl's catalog, you'll have an unacceptable collision rate.
Hashes are attempts at creating a _truncated_ "unique" representation of an input. They throw away data they hope isn't necessary to uniquely identify between possible inputs (bits). A perfect hash for all possible 32 bit values is 32 bits. You can't even have a collision free 31 bit hash.
So back to the blurry security camera footage of a license plate or a face. Sure, that "hash" can reliably tell you that it wasn't a sasquatch that committed the robbery, but it literally doesn't contain the data necessary to _ever_ prove it was the suspect in question, even if the techs _can_ prove that the suspect hashes to the image in the footage.
Crafting a perfect hash function with keys being the set of words from the OED is perfectly reasonable. It’ll take a short while to produce it, but it’ll work just fine. (rust-phf says that it “can generate a 100,000 entry map in roughly .4 seconds when compiling with optimizations”, and the OED word count is in the hundreds of thousands.)
Thanks for the rust-phf link. I'm bookmarking for my next project!
For a face, sure, for printed text/license plates there are effective deblurring algorithms that in some cases may rebuild a readable image.
A (IMHO good) software is this one (was freeware, now it is Commercial, this is the last freeware version):
https://github.com/Y-Vladimir/SmartDeblur/downloads
You can try it (just for the fun of it) on these two images:
https://articles.forensicfocus.com/2014/10/08/can-you-get-th...
https://forensicfocus.files.wordpress.com/2014/09/out-of-foc...
https://forensicfocus.files.wordpress.com/2014/09/moving-car...
For the first choose "Out of Focus Blur" and play with the values, you should get a decent image at roughly Radius 8, Smooth 40%, Correction Strength 0%, Edge Feather 10%
For the second choose "motion Blur" and play with the values, you should get a decent image at roughly Length 14, Angle 34, Smooth 50%,
If you want ease collision examples you can take a look at people using CRC32 as hashes/digests. It is notoriously prone to collisions (since only 32 bits).
You can't compress a file by repeatedly storing a series of hashes, then hashes of those hashes, down into smaller and smaller representations. The reason that you cannot do this is that you cannot create a lossless file smaller than the original entropy. If you could happen to do so, however, you would get down to ever smaller files, until you had one byte left. But, you could never decompress such a file, because there is no single correct interpretation of such a decompression. In other words, your decompression is not the original file.
But in real life there's collisions.
And in real life image or sound compression, blurs, artifacts and resolutions, it is fundamentally destroying information in practice. It is no longer the comparatively difficult but theoretically possible task of reversing a perfect hash, but more like mapping a name to the characters/bucket RXXHXXXX where x could be anything.
There are lots of values we can replace X with which are plausible, but without an outside source of information, we can't know what the real values in the original name was.
It went from this:
https://media.giphy.com/media/pUf3YfamV7BV6/giphy.gif
To this:
http://img.go-here.nl/Roundhay_Garden_Scene.gif
The funniest part was that the resolution really goes up if you make 1 px into 40 and align the frames accurately (then adjust opacity to the level of blur)
The crime television thing would be possible if you have enough frames of the gangster.
To play devil's advocate though, modern neuroscience and neuropsychology basically tells us that that our brains reconstruct and recreate our memories every time we try to remember them. Our memories are highly malleable and prone to false implantation... and yet witness testimony is still the gold standard in courts.
http://www.cs.cmu.edu/~aayushb/pixelNN/freq_analysis.png
The computer is fantasizing.
This is another avenue that could be further explored, which I quite like. That is, a non-artist can doodle images and create a completely new photo-realistic image based on the line drawings.
I was modifying a few images (from link on another comment here: https://affinelayer.com/pixsrv/ ) and the end results were interesting.
It could also be used to generate a false confession. If the prosecutor says "We have proof you were there at the scene" and shows you some generated image, then you as an innocent person have to weigh the chances of the jury being fooled by the image (and even if it's not admissable in court, it may be enough to convince the investiging team that you are responsible and stop looking for the real perpetrator) and the expected sentences if you maintain your innocence vs "admitting" your guilt.
Although what we don't have is any certainty that the enhanced face actually looks like the killer.
If 2017 was the year of GANs, 2018 will be the year context prediction.
This would of course not give you any space savings when you want to distribute 1 movie. There would be some minimum number of movies where the training set + actual movies would be smaller than the sum of the sizes of the individual movies compressed.
I'm not saying this would be a net space saver, or necessarily a good technique at all, but the concept is intriguing.
How can their algorithm work out the skin tone from a colourless image. Perhaps their training data only had white people in it?
> I don't understand how the edges-to-faces can possibly work. The inputs seem to be black & white, and yet the output pictures have light skin tones.
The step you're missing is that an edge detector is run on the entire database of training images to produce a database of edge images. The input edge image is run against that corpus of edge images in order to find which edge images match, then sample the corresponding original color images and synthesize a new color image.
But shoes and cats are one thing; reconstructing people's faces is another. I know the paper & the authors are demonstrating a technology here, rather than directly saying "you can use this technology for purpose X", but the discussion in these comments has jumped straight into enhancing images and improving existing pictures/video. But there is a very big line between 'reconstituting' or 'reconstructing' an image and 'synthesising' or 'creating' an image, and it appears many people are blurring the two together. Again, in the authors' defence, they are clear that they talk about the 'synthesis' of images, but the difference is critical.
That's right. But with the caveat that a large training set can determine plausible colors and rule out implausible ones. This is more true for faces than for shoes! The point is that there is some correlation between shape and color in real life. The color comes from the context in the training set. This is what @cbr meant nearby re: "skin color is relatively predictable from facial features (ex: nose width), it should be able to do reasonably well."
There are CNNs trained to color images, and they do pretty well from training context: http://richzhang.github.io/colorization/
> there is a very big line between 'reconstituting' or 'reconstructing' an image and 'synthesising' or 'creating' an image, and it appears many people are blurring the two together.
Yep, exactly! Synthesis != enhance.
Just imagine the kind of CCTV usage being discussed elsewhere in this thread. But the neural network happens to have a wrong bias towards skin colour...
One of the best articles I've read on the topic, if you're interested: https://medium.com/@blaisea/physiognomys-new-clothes-f2d4b59...
I wonder if there could be a way to "index" those images so I can find them back without storing the whole image, using some type of clever image histogram or hashing-kind function.
I wonder if that thing already exist, since there are many images, and since most images have a lot of difference in their data, could it be possible to create some kind of function that describe an image in a way that entering such histogram redirects to (or the closest) the image it indexed? I guess I'm lacking the math, but it sounds like some "averaging" hashing function.
This is actually training a neural network on the Markov model, so it's very similar to core ideas behind the OP's paper. The core idea is to model the probability of a bit of sound by breaking it into the last note and everything that comes before the last note ("P(audio)=P(audio∣note)P(note)"). If you sample a bunch of audio and factor it that way for any given point in time, and accumulate that data somewhere, you can then sample the accumulated data randomly to generate new music.
There are other audio NN synthesis methods as well, pretty sure I've even seen one posted to ShowHN before.
For example an approach similar to https://gauthamzz.github.io/2017/09/23/AudioStyleTransfer/ but then using the methods described in the PixelNN paper.
It more or less attempts to be what you describe. Not very polished yet, but I had some basic success in modeling the parameter space of a synth, and adding new latent spaces with regularization.
(Eg first result: http://sunbeam.ece.wisc.edu/csaudio/)
I do wish they would release the code AND any related training images they used to get those results.
EDIT: Furthermore the range of plausible imaginary images that match a given input is high (infinite?).
This paper does not demonstrate an enhancement technique but a phenomena which those using inverse methods called "overfitting".
(Because these kind of techniques aren't really enhancing the images in a way that gives you new and useful information: they are taking the low-res images as input, and giving you a plausible high-res image as output, based on it's training data. It is NOT however trying to say "this is the ACTUAL high res image that generated this low-res image"
It's not that far off actually, but they are talking about nearest neighbor Markov chains, not interpolation. You probably already know nearest neighbor Markov chains because there are lots of text examples, and a ton of Twitter bots that are generating random text this way. The famous historical example was the usenet post that said "I spent an interesting evening recently with a grain of salt." https://en.m.wikipedia.org/wiki/Mark_V._Shaney
This paper does use a NN to synthesize an image, which is conceptually pretty simple, even if difficult to implement well. After that they use a nearest neighbor Markov chain to fill in high frequencies. The first paper referenced is also the simplest example: http://graphics.cs.cmu.edu/people/efros/research/EfrosLeung....
That paper fills missing parts of an image using a single example, by using a Markov chain built on the nearest neighboring pixels. That paper is also one of the only image synthesis papers (or perhaps the only paper) that can synthesize readable text from an image of text. That's really cool because the inspiration was text-based Markov chains.
Well, of course almost anything can be interpreted as a Markov process, but I don't think it's a very useful abstraction here.
Oh, it absolutely does. I think it's fair to say that Efros launched the field of nearest neighbor texture synthesis, and his abstract states: "The texture synthesis process grows a new image outward from an initial seed, one pixel at a time. A Markov random field model is assumed, and the conditional distribution of a pixel given all its neighbors synthesized so far is estimated by querying the sample image and finding all similar neighborhoods.
This is the same Markov model that all subsequent texture synthesis papers are implicitly using, including the paper at the top of this thread. Efros' paper implemented directly is really slow, so a huge number of subsequent papers use the same conceptual framework, and are only adding methods for making the method performant and practical. (Sometimes, at the cost of some quality -- many cannot synthesize text, for example.)
Note the inspiration for text synthesis, Shannon's paper, also describes the "Markoff Process" explicitly. http://math.harvard.edu/~ctm/home/text/others/shannon/entrop... (Efros referenced Shannon, and noted on his web page: "Special thanks goes to Prof. Joe Zachary who taught my undergrad data structures course and had us implement Shannon's text synthesis program which was the inspiration for this project.")
> Well, of course almost anything can be interpreted as a Markov process, I don't think it's a very useful abstraction here.
It's not an abstraction to build a conditional probability table and then sample from it repeatedly to synthesize a new output. That's what a Markov process is, and that's what the paper posted here is doing. I don't really understand why you feel it's distant and abstract, but if you want to elaborate, I am willing to listen!
And I didn't mean that Markov processes are abstract in any "distant" sense, but that they are an abstraction, ie a "perspective" from which to approach and formulate the problem.
"To address these limitations, we appeal to a classic learning architecture that can naturally allow for multiple outputs and user-control: non-parametric models, or nearest-neighbors (NN). Though quite a classic approach [11, 15, 20, 24], it has largely been abandoned in recent history with the advent of deep architectures. Intuitively, NN works by requiring a large training set of pairs of (incomplete inputs, high-quality outputs), and works by simply matching the an incomplete query to the training set and returning the corresponding output. This trivially generalizes to multiple outputs through K-NN and allows for intuitive user control through on-the-fly modification of the training set..."
Note the first reference #11 is Efros' non-parametric sampling, and that the authors state this is the "classic approach" that they apply here.
What you call "interpolate every pixel independently from its surroundings" could be another way to describe a Markov chain, because 1: it is sampled according to the conditional probability distribution (which is what you get by using the K nearest matches.) and 2: the process is repeated - one pixel (or patch) is added using the best match, then it becomes part of the neighborhood in the search for the pixel/patch next door. The name for that is "Markov process", or in the discrete case, "Markov chain", if you take an unbiased random sample from the conditional distribution. If you always choose the best sample, then it's the same as a Markov chain, but biased.
> (neighbor means a close-by pixel in the training set in the feature space, not a spatially close pixel)
That's right, and that's why it's misleading to talk about nearest neighbor interpolation, because that phrase is a graphics phrase that means interpolate from spatially close pixels. Hardly anyone else calls it interpolation, they call it sampling, point sampling, and other terms.
*EDIT:
I'm going to relax a little bit on this. "Non-parametric sampling" is a tiny bit different from a Markov process in that a Markov process attempts to simulate a distribution in an unbiased way. By using the best match instead of a random sample from the conditional distribution, the output may produce a biased version of the original distribution. This is why it's called non-parametric sampling instead of calling it a Markov chain, but the distinction is pretty small and subtle -- texture synthesis using non parametric sampling is extremely similar to a Markov chain, but not necessarily exactly the same.
Side note, it's really unfortunate they used the abbreviation "NN" to talk about "nearest neighbor" in a paper that also builds on "neural networks".
It seems that this could be used in theory with any feature descriptors, such as local color histograms, although the results wouldn't probably be as good.
Edit: Being a nearest neighbor probably also carries the usual computational complexity problems of the method. If I understand it correctly, they ease this by actually first finding just subset of best matching full images using the CNN features and then do a local nearest neighbor search just in those images.
https://en.wikipedia.org/wiki/Nearest-neighbor_interpolation
versus
FWIW, what you just described is known as a "Markov process". It is sampling a known conditional probability distribution.
While some interpolation of the data happens because the output represents a mixture of the training images, this is not "interpolation" at the pixel level, it's picking best matches from a search space of image fragments. (And the pixel neighbors are usually synthesized - the best match depends on previous best matches!) This is distinctly different from the kind of nearest neighbor interpolation you'd do when resizing an image.
Note the phrase "nearest neighbor" in this paper has an overloaded double meaning. It is referring both to pixel neighbors and neighbors in the search space of images. The pixel neighbors provide spatial locality within a single image; this is how & why high frequencies are generated from the training set. Nearest neighbor is also referring to the neighborhood matches in the search space, the K nearest neighbors of a given pixel neighborhood are used to generate the next K pixel outputs in the synthesis phase.
Take a low resolution input image, and hallucinate a higher resolution version by statistically assembling bits from similar images in a large data set of training images.
Why is that? Blame Kolmogorov. There are deep connections between compression, serialization, and computation. An optimal compression scheme is a serialization and the Turing-complete program to decode it. For example: you can compress pi into a few lines of algorithm plus a starting constant like 4.