5 ms·
Really? Seems blurred enough to me that even some sort of ML would spit out wrong characters.
by boppo1 5y ago
Really? Seems blurred enough to me that even some sort of ML would spit out wrong characters.
- birdyrooster 5y agoNot at all, I can already make out the characters and recognized quickly that all but the last 4 blurred characters are hexadecimal and the last appear to be [a-z][A-Z]
- seba_dos1 5y agoIt's blurred lightly enough for me to be able to tell at least some characters just with my eyes.
- generalizations 5y agoKnown font, known range of possible characters (almost certainly ascii), and probably several blurred characters in there that we know (like the t in attachment). If the blur is differentiable per-character, it's probably reversible.
- warent 5y agoEven without automation. Just make a list of blurred alphanumerics, then match every blurred character in the picture against the character list. with some patience probably doable in a single day
- spicybright 5y agoProbably the most time efficient way to do it tbh.
- ansgri 5y agoThat’s called generative recognition, and it’s the best way to do it (probably can be formally proved under some noise model assumptions) if you have a good model of the data and enough computational resources.
- boppo1 5y agoWhat tools do people use for stuff like this? A python library? C++?
- derac 5y agoYou don't need ml. If you know the font and blurring algorithm (or a close approximation), you blur all letters of the font at that size and compare output.
- dan-robertson 5y agoMost of the time, blurring is down with a Gaussian blur and this is, in theory, reversible. In practice what is actually done is round_to_fraction_over_256(gaussian_blur(point, image, …)) so there is an error in the unblurring process. But this error is often not insurmountable, especially with extra information.
- lmilcin 5y agoActually, if you can identify the exact font and exact location of each letter (as is in this case) it doesn't matter what kind of blur it is. If you assume the photo was made with one of the consumer applications, there are only so many popular bluring algorithms. You can brute force it quite easily by testing each character and each type of blur until you get exact match (and in many cases even inexact will suffice).