| General > General Technical Chat |
| Can ordinary humans enhance blurry pictures yet ? |
| (1/8) > >> |
| MathWizard:
I'm working on a repair , and I lost some of the notes I took when I was making a schematic. But I found some old picture of the PCB on the web. And I found a missing resistor not in my notes, but from the picture, I can just make out the resistor is actually just a jumper. But that reminds me, since the 90's as a kid, in all kinds of TV and movies, people are always making blurry pictures look better. Are we there yet ? I don't see any button in windows 10 picture viewer for it, or is some math for it actually built in?. Or can I get a freeware program for this yet ??? |
| MikeK:
No, and we never will be. There is no information to extract from a single pixel. Alternatively, just say "Enhance!" |
| Gyro:
Ah, the old Hollywood zoom into a tiny fragment of an image and enhance. ;D You might try importing the image into something like Gimp and trying a little sharpening and contrast enhancement, but you're not going to be able to reveal detail that isn't there, just pep it up a bit. |
| T3sl4co1l:
There are plenty of ways to enhance resolution -- but doing so is increasingly questionable, statistically speaking. We can use an L1-norm interpolation for example, which gives lower noise results from surprisingly sparse inputs; we can also use neural networks to fill in what's likely missing, but this is contingent on the data the neural network was trained on -- that is, if it was trained on pictures of buildings, it's more likely to find stucco or brick or wood textures (or some hallucinogenic blend of everything!) when zoomed in, regardless of what the actual texture was (e.g., maybe it's actually a smooth metal surface, but the optical noise of the camera is being interpreted as real texture). The worst outcome is trying to speculate on something that well and truly has no information to support it. We could take a ten pixel image and expand it full screen. Obviously, the number of textures, and objects, that can be contained in that full size image, vastly exceeds that of the source image -- and just about any of them are as statistically likely as another, given that the overall colors are approximately right to match the original image when this image is reduced. Today, we can very easily create images that are a shockingly similar to a speculative organic process, as if using ones' imagination to fill in the blanks. It's like a dubious witness, testifying of a story, which they seemingly created from whole cloth; whereas the scene they were actually witness to, was perhaps a few seconds of shouting. And similarly, if we are to put too much trust in such a representation, we may find we are creating some very bad outcomes indeed (e.g. wrongfully imprisoning someone on false evidence). Okay, so going from basic image processing, to legal proceedings, is a rather abstract step, I admit; but to the extent that photographic evidence might be used in such proceedings, it's very applicable. And more broadly, we can guess that analogous processes occur within a witness's mind -- that is what I'm getting at. So -- if your image isn't very blurry, maybe an extra 20 to 100% of resolution could be drawn from it, maybe even 200 or 400% in extreme cases; but fine details will very quickly become questionable at that rate, especially if the algorithm is prone to synthesizing meaningful details (like text, or faces) where there isn't necessarily anything of the sort to begin with. As for actually trying it to see, I'm not sure offhand where you can find such a processor? There may be some online tools, worth looking around to see. I'm sure Photoshop has all manner of clever filters available. Or you could download some of the tools these are based on, which may require some degree of licensing, or building (e.g. from source code, or the tools just aren't very convenient e.g. CLI based), or cooking up utterly from scratch (you can do all manner of machine learning yourself, for free; but, it is an entire field to learn about, good luck -- and truly, good luck, it will surely prove very useful/profitable to put the time into, if you choose to!). Tim |
| coppercone2:
replaced with something a little more PC |
| Navigation |
| Message Index |
| Next page |