4 ms·

I can barely count how many lines the text message has, and you think a ML tool would be able to decipher it? This is an impossible task, the text information i

• 0xfaded 5y ago
  A gaussian blur is a point-wise multiplication in Fourier space. If you know the filter, you can technically reverse it with a point-wise division. Even noise can be controlled for by using optimization techniques to recover the correct divisor. It doesn't at all surprise me that you could train up an ML model to recover the original.
  • Etherlord87 5y ago
    I'm not saying it's impossible, but remember that information is lost in the blur. It's as if you posterized an image to two colors - just because you know the algorithm doesn't mean you can revert the process.
    • aliswe 5y ago
      This is an induction based argument though. Because information is lost (is it?), it's probably impossible to reverse. Look at this example image,do you think it would be possible to restore? https://kbpdfstudio.qoppa.com/wp-content/uploads/sample-text-deblur.jpg https://kbpdfstudio.qoppa.com/wp-content/uploads/sample-text... Also this depixelated text: https://www.hackread.com/wp-content/uploads/2022/02/new-underactor-tool-reveal-pixelated-text-expose-sensitive-data.jpg https://www.hackread.com/wp-content/uploads/2022/02/new-unde...
      • Etherlord87 5y ago
        I think I made a mistake by saying "information is lost", rather than "some information is lost" - clearly, some information remains, and as I started my comment, I'm not saying it's impossible. In fact I share the sentiment that seemingly irreversible things can often be reversed, we just don't have the technology to do so. However my enthusiasm took a huge hit once I learned how the technology we have usually operates: for example the photo sharpeners enhance photos by creating faces different than the actual original pre-blur faces. Just got an idea I have to test eventually, how close would those "recovered" images be to the algorithm input, if the blur was reapplied... Your first example (recovering text from photo) is impressive, but the blur is clearly one that blurs as little as possible to make the text unreadable. I think if you increased the blur just slightly, too much information would get lost. I also wonder what was the resolution of the image and what was the color depth (was that a JPEG or RAW format?) Without the ability to inspect and repeat the process I'll remain skeptical about de-blurring.
        • aliswe 5y ago
          If blur is a strictly mathematical and deterministic operation, which it often is, then reversing it should be possible. Remember that a pixel is only a bunch of numbers. Even non-mathematic blur should be possible... And the reason for that is small variations in the blur still exist and give away (with varying amounts of certainty) which shapes could be put together to form the resulting blurry pattern. But yes, information is actually lost in the sense that more than one image can potentially be the used to generate the same blurry pattern, and there is also limited precision to each pixel.
          • theWreckluse 5y ago
            Just floating the point: Not all mathematical, deterministic operations are invertible.
• antifa 5y ago
  I think it's still prudent to move the task from "probably impossible" to "definitely impossible".