Reliable OCR software?

[Pocokk]

Hey there,

Been wanted to ask it ages ago, but never got the chance: which OCR software would you wholeheartedly recommend that can reliable work on not so clean, scanned documents, ebooks, pdfs? I'm planning to delve into some Letter Pages on ancient (50-60 years old) scanned comic books, however I'm not getting a whole lot of success with various software: switching characters, seeing dirt on the page as a symbol, can't differentiate between 8 ( S characters etc.

Are there good free alternatives or at that point a purchasable, high end option is the only good choice?

What do you guys think?

Thank you in advance!

[GitHubRulesOK]

I first sold OCR 4 Decades Ago for recognising single drawn or written symbols and OCR is roughly as good now as it was then. That is it needs intelligence to spot the failures. So today we have infant like AI that can improve over the next 40 years much much faster.

Ghostscript and MuPDF which is the base of SumatraPDF can do fairly well so compare the old image test I re ran this week via SumatraPDF and copy the characters to see how they compare. But the contrast between foreground and background will always need to be cleaned up whatever the thresholding system used.
<img width="247" height="147" alt="image" src="https://github.com/user-attachments/assets/ae398e5f-115f-4c4b-81c2-8851b2a9939b" /

Note your browser (Edge) is often the best as using online AI so is able to compensate by looking across the web for fuzzy character patterns

 PDF" (also known as image-only PDFs

image-mupdf-ocr216.pdf

[Pocokk]

I tried out Ghostscript and the result is not very satisfying, see attached pdf. That's the type of "dirty scanned file" I was talking about. If it were a brand new digital scan, I imagine it would give you a very nice document at the end, but that's not the case here.
output_ocr.pdf

[GitHubRulesOK]

No "Ordinary" OCR will be that good as the page has mixed qualities thus needs compartmenting into ROI areas so then each area can be treated by AI across a large corpus and thus produce results the cost of say buying Microsoft.

and even then its a FAIL if you use your Own Eyes to compare OCR (One at a time Character Recognition.)

Thanks to Ed Sale's amazing Secret System,
you can play a lovely song the first day.
any song by car of note in 7 days. In this
SPECIAL INTRODUCTORY OFFER
you get all this: 66-pase Intruction
Book, 52 photos. 87 chord and Anger
placing charis. 110 songs (words and
music), chord finder of all popular
chords. Special Guitarist's Book of
Knowledge, and extra bonus:
special wallet - size tuning
aid for tuning any quitar by ear.
Total Value $9.00


Add $1.00 shipping & hndig.
Money back Guarantee
ED SALE, Studio 13J. Grand Island, Fle. 32735A
```

[Pocokk]

Yes it is very tricky, as these are non standardized formats, none of the programs can reliably OCR all of them at once and even then it makes countless error, seeing strings where there are none, can't differentiate between certain characters or mix-matching them up. It's a struggle, really.

Another example below, which is a much cleaner scanned image, yet it still need a lot of touch-ups afterwards. I guess it won't be an Easy OCR (process), like ever!

[GitHubRulesOK]

Most OCR will need analysis and possibly some pre-processing which is what AI is good for then there is a chance that slowly an OCR app can run optimised so here I saved your image after a little tuning in a graphics app and then saved as PDF and then one click takes a while to convert PDF to OCR.PDF . However WITHOUT settings for AREAS in order the output will be simply jumbled even if there were no errors.

This is where OCR apps like Abbyy can often fare better than basic OCR but for high-speed separate into columns etc then it needs Computer Vision RULES to follow. the bottom line is it used to cost the price of a house to get close but never perfect results.