6
$\begingroup$

So let's say you had a really nice day in a flight simulator and you are getting videos of this type of quality:

A snapshot of the compressed video

This is Full HD (1080p), but heavily compressed. You can literally see the pixels. Now I tried to use something like RAISR, and this python implementation, but it only scales the image up and does not 'fix the thicc pixels'.

So is there a type of AI that does fix this kind of video/photo into a reasonable quality video? I just want to get rid of those pixels and image artefacts that was generated during the compression.

$\endgroup$
  • 1
    $\begingroup$ Although there are AI techniques that can do this sort of thing, this Stack Exchange is not for questions looking for implementations. It is for questions about the nuts and bolts or how they work, or even the philosophy of whether or not we could or should build "intelligent" machines. If you are willing to learn enough about neural networks and invest a few weeks or months into researching and writing an algorithm to de-pixelate images, then you are in the right place. If you just want to fix your image, then sorry no - I believe there are projects out there that might help though. $\endgroup$ – Neil Slater Oct 14 '18 at 19:58
  • $\begingroup$ @NeilSlater I am always open to design new algorithms. I just wondered if there was already such algorithm out there. $\endgroup$ – Hugo Woesthuis Oct 15 '18 at 7:21
  • 1
    $\begingroup$ Yes the algorithm is out there, it is not much different to super-resolution, and implementations are available if you search. At one point I am pretty sure I have used a NN trained to de-block an image on my own graphics. If you are interested in studying the algorithm, could you please make that clearer in the question, explaining what you know already? $\endgroup$ – Neil Slater Oct 15 '18 at 7:24
  • 3
    $\begingroup$ Here's one example of an algorithm built to do what you want: arxiv.org/pdf/1709.06229.pdf $\endgroup$ – Neil Slater Oct 15 '18 at 7:32
  • $\begingroup$ @NeilSlater Very interesting! Never heard of it lol. $\endgroup$ – Hugo Woesthuis Oct 15 '18 at 7:45
1
$\begingroup$

Check out source code to DeepImagePrior it does a remarkable job guessing what's missing to repair images with a variety of damage.

$\endgroup$
0
$\begingroup$

I've answered a part of the question below, i.e. can the photo/video be decompressed in theory?

If the RGB values are continuous (well behaved function), interpolation should generate the pixels in between them in a predictable pattern like say in old pictures of models where skin color shades and data is available in plenty for models today and they resemble each other whereas in outlier cases like this picture where the designs of cockpit circular meter maybe different for each plane or maybe too much data is lost say of a distant object, no neural network should be able to fix it because the possibilities of an orange blur can be an orange, a sun or many other things.

One of the solutions to such things can be recognizing context but finding sufficiently dense pixels so exact should be difficult (I think impossible) because sun will be very bright but orange might have some spots.

In short, average things can be done, not outliers.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.