I have an image that is really blurred and I wanted to know if there were any method to get the text of a low quality picture ?

enter image description here

It comes from this link. This is a page from this newspaper online.


A simple, effective and ethical method exists to get the text from THAT specific low quality image: pay the publisher the fee.

  • 1
    Wish I could up vote this many times. – Scott Feb 6 '19 at 2:15
  • And if you politically oppose to this publisher, would you still pay him the 50$ he asks for his image? – Revolucion for Monica Apr 4 '19 at 17:33
  • Though I might disagree politically with that publisher... or any publisher... nonetheless they retain their IP rights - if simply disagreeing with someone were to somehow nullify their IP rights, then what value is the very idea? If for example your use-case is sheer politics, or satire, this may be permissible "fair use" in which case my objection disappears. If however, your use involves making a profit, or creating a non-satirical derivative work for commercial use, then you pay the one-time fee and get your task done. – GerardFalla Apr 4 '19 at 17:59

There is nothing you can extract form that image. Too few pixel to work with sorry. Real life isn't like CSI or Blade runner.

  • 1
    "Enhance." clickity-click. "Enhance." clickity-click. – 13ruce Feb 5 '19 at 21:53

It's not only low quality, I think it's intentionally selected low resolution to make possible to see the general layout, but not to read the content. So, if you find a way to read it you should also pay to the publisher what he wants. He obviously has stated a price for reading it.

No sharpening cannot restore the text, but the recovery can be based on the fact that it's text and has a certain font and size. In addition the text is written in certain language. That should give some basement to guess which words has been written. I can give a simple example:

enter image description here

On the top I have two words written in not so high resolution. Those words are Auto and Home.

BTW. Just in case you need higher resolution to recognize common fonts, the used font is Bodoni MT.

The three darker patterns are comparisons with blending mode difference. Auto vs Auto and Home Vs Home of course both give difference = 0, but the average pixel brightness difference is about 50 (or as well 20%) in Auto Vs Home. Maximum would be 255, if there was a negative.

Actually computer vision does the same, but with more developed calculation algorithms. It makes comparisons with all known detectables and decides the most probable cause of the seen blurred patterns. It isn't only in movies or crime investigation tv series, but in real life. For ex. the security provider of a jewelry shop is alarmed as soon as the face recognition machine sees some well known (=police familiar) person walking around. In the same way the police recognizes interesting (=wanted or non-existent) car license plates routinely.

The applications are numerous in science, business, intelligence, surveillance, robotics, weaponry, probably more various than I can imagine. Of course the cost also is astronomic, but those who want the best do not start by asking the price.

Unfortunately I have no guidance to give how ordinary people could get such sophisticated pattern matching software nor does it be capable enough for your case. If you don't find it or it happens to be too ineffective, you can use the alternative reading method which already is suggested by others.

  • 1
    Number-plate recognition is ubiquitous these days - even supermarkets have it to see how long you've been in their car park & the police/DVLA use it to see if your car is taxed & insured, millions of cars per day, travelling at motorway speeds. That's not really the same as reconstructing text from the tiniest blurred image. – Tetsujin Feb 5 '19 at 19:11
  • @Tetsujin maybe the needed computation is lighter for car number plates, but definitely it must be based on pattern matching. – user287001 Feb 5 '19 at 19:16
  • maybe so - but it's based on starting with an image known to be sufficient for accurate recognition, rather than a random tiny pic. Inaccurate recognition would be a 'national disaster' as this country no longer has any other form of taxation notification physically attached to the vehicle, it is done entirely by remote recognition cameras, millions of them around the country, like speed cameras [though not the same devices]. – Tetsujin Feb 5 '19 at 19:18
  • Pattern matching or not, blurred or not. You can not extract info below the nyquist rate. – joojaa Feb 6 '19 at 5:05
  • @jooja Not true! The info symbol rate can be radically lower, so aliasing the carrier signal doesn't necessarily remove it. An exact constructive proof: Let's have a sinewave, say 1 kHz, Then let's have an alternative signal = full silence. Let the transmitted signal be allways one of these and the signal sender selects which is valid in each second GMT. Let the signal be sampled with Fs=900Hz, well below the Nyquist rate. I believe you could still be able to decide which one has been sent if you know the alternatives. – user287001 Feb 6 '19 at 8:57

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.