Machine learning is really machine-enhanced educated-guesswork, which has its place but also has its limits.
Being able to read the title on the books in the example photo is great; you could rely on the title for evidentiary purposes, the smaller text probably not so much. So for a security camera it would do poorly at identifying the color of a car, but might well be sufficient to read the license plate.
You show in a courtroom a CNN-enhanced low light image of a car and it's there, literally 'clear as day' - the jury will find it pretty compelling. But maybe the data really wasn't there in the original image, and the CNN just filled in some blanks based on previous images of license plates, letters, and just random noise it had seen in the past.
The worry is when these kinds of algorithms get built in to basic image capture processes, so you never even see the raw data, only data that has already been filtered through the inbuilt prejudices of the CNN enhancement suite.
The camera never lies, but now it doesn't have to, because it can convince itself it saw something that wasn't really there...
You are right that the raw sensor data should always be preserved. But sticking with the license plate example, you could challenge a picture of a single car with a visible license plate far away in a wooded area, but it would be hard to refute a picture of the same car in a parking lot surrounded by other (non-suspect) vehicles whose presence there at the same time could be independently verified. In other words, if I can show that it accurately read the license plates of 9 other cars, the chances that it got yours wrong go way down.
That's assuming a single photo taken in the dark by an investigator. With a fixed security camera you would have an even larger basis of comparison, with a population of hundreds or thousands of license plates against which to rate it. I predict that before long we'll see preemptive certification for devices warranting the reliability of their image pipeline out to a certain distance at either the manufacturing or installation stage.
https://www.theregister.co.uk/2013/08/06/xerox_copier_flaw_m...
License plates are an ideal breeding ground for false enhancement owing to standardisation of appearance; an ML algo trained on lots of examples might, without due care, learn to replace as a well-known texture.
Also, y'all need to think more like prosecutors. Say you are dragged you into court on the basis of photos showing your car in the dark, and you object that the photo is from the ML 9000 security camera and it might be just imaging your license plate. The police/prosecutors will just 'borrow' your car and leave it there for a night and leave it up to the jury.
Forensic evidence can be and is regularly abused, but it can also be quite easily validated and it's massively persuasive to juries.
Let’s assume I’m innocent but some neural net has placed my car at the location of a crime.
You’re saying that if I challenge the evidence, the prosecutors will counter that by showing that if my car were there, the neural net would have produced a picture of my car? They don’t need to do that and it adds nothing to their argument. I’m not challenging that the neural net is capable of producing an image of my car.
No, the point is I am placed in the position of having to demonstrate that there exists some other car which under those lighting conditions the neural net would mistake for mine. That’s a far harder burden of proof for me to reach.
Honestly this is similar to the way fingerprint, DNA and hair sample matches are presented to courtrooms all the time so it isn’t a new problem. As you say, forensics are persuasive.
For a long time, photographs were typically used as records. Even when they were art, they were typically records of something. Soon, we'll be typically doing so much more with them and will have to accept that photos can, but frequently won't serve as records.
There is an entropy limit to the message, but the message isn't actually the only data.
One thing humans are great at is integrating existing knowledge into a messy situation and intuiting more than is available just from the raw message.
I.e. The message has an entropy limit, but the message isn't the whole dataset.
First of all, if you were to use the image as a communication channel, how much you could theoretically communicate is exactly the entropy of messages (by definition), and optimal communication means maximum entropy. Information theory already assumes shared existing knowledge in the form of codes; the codes essentially encode all this knowledge, which you could make an analogy in images to digits and shapes, etc. -- what makes them decodable is there is statistical redundancy, shapes do not occur arbitrarily (i.e. not every possible shape occurs, at least not with equal probability) and exhibit dependence between different pixels of a shape and even between shapes elsewhere in the image. Again the dependence is (generally) based on the statistics of the distribution of all possible images -- it essentially encodes all prior knowledge.
This redundancy allows reconstruction of losses in on part of the image from data elsewhere. It's the same principle used in error correcting codes, except codes are designed, while shapes are mostly natural (except things like alphabets, which are designed and indeed follow some principles of codes). But because they're not designed there's not guarantee of having a unique/reliable decoding (i.e. you can get a distribution).
I think that's a significant issue, because if your estimate doesn't match reality it could have important consequences for the use of the image: perhaps text goes from 'X is good' to 'X sucks'.
In this case a few things could be done:
1) Have some kind of watermark indicating the image was enhanced by a neural network, and possibly contains false information;
2) Have some kind of indication of reliability of the image: it should encode the multimodality/confidence of the decoding distribution -- how many different solutions does this have. If it is more or less unique, it would show as high confidence; otherwise it would show a low confidence indicator;
3) Instead of trying to convey uncertainty, the system could simply give up in cases where there is too much uncertainty, i.e. leaving the image dark. This could be done locally or globally, although locally it introduces a lightning consistency problem.
---
There's another important observation w.r.t. Information Theory/Statistics: it essentially assumes unbounded computational power (since this distribution could require analyzing arbitrarily large datasets). Of course this isn't true in reality. For example, the entropy of an encrypted of a redundant text is exactly the entropy of the plaintext string plus the entropy of the key (given an encryption ensamble or encryption prior) -- the process of (e.g. through brute force) finding the key doesn't concern statistics. However, with reasonable computational power, the (properly) encrypted stream is indistinguishable from a random string, hence it would have maximum entropy. So there are further computational limits beside statistical limits. In the case of encryption the function is again designed (to be not tractable), while in natural images the correlations are of simpler and hopefully more tractable nature (although I'm sure not always the case).
Eyewitness testimony is awful but given gold status.
It's not trying to make things readable; it's trying to make things look like there was more light in the room when they were shot. In rooms with high lighting, some objects have glare. That's "correct"—it's what would appear in the training data.
Wait, did it? Isn't the middle photo being shown for comparison only, rather than as an input?
EDIT: Finally got the paper to load via the helpful wayback machine link provided in another thread. It looks like the goal was to simulate a long exposure with a short exposure. So whitewashing of "bright" areas in the original might be expected.
You do realize that, in a camera sensor, light is the signal, right? So the more light, the higher the signal-to-noise ratio, which means that yes it does have more information available to extract.
And yes, I quite realize that the input is (a). I'm guessing that in your display you are not seeing that there is a brighter spot in the middle of (a) corresponding to the whitewashed area in (c). Try maxing out your brightness if you're on a phone or laptop and you should see it. I can even make out letter shapes in (a) within this bright spot.
people bring this up all the time as hot-takes in these areas. it's conditional inference. it's no more disingenuous than linear regression.