Hacker News Reader

Comment by atonse | Hacker News Reader

I must’ve done things spectacularly wrong cuz the two times I’ve tried tesseract (second time to recognize factory printed 8 inch tall letters on a trash can), I got 0% accuracy. 0%. Not even close. So I gave up.

View on HN

Yeah tesseract is more like the bit that would sit in the middle of an ocr solution than a complete solution. But it’s all we’ve got for free at the moment.

You pretty much need black text on white background at 300-600 dpi. (Not sure the exact size but I’ve had crappy scans do better by scaling the file.)

I’ve had reasonable success with photos of printed pages run through text cleaner.

> But it’s all we’ve got for free at the moment.

Doesn't OCRopus qualify as well (it does look unmaintaned, or less actively maintained than Tesseract)?

You typically need to pre-process the images.

I'd recommend https://scantailor.org/ for this (OSS, but unmaintained)

This has the latest developments, but is also seemingly unmaintained for over a year: https://github.com/4lex4/scantailor-advanced

Scan Tailor forum: https://forum.diybookscanner.org/viewforum.php?f=21

> [ScanTailor Advanced] seemingly unmaintained for over a year

ScanTailor official repo also archived on November 29, 2020.[0]

> This project is no longer maintained, and has not been maintained for a while.

As alternative to ScanTailor (and its forks) there is gImageReader (Tesseract Qt/GTK GUI), but also seems like unmaintained since 2019.[1,2]

[0] https://github.com/scantailor/scantailor/commit/e881b30b6ed1...

[1] https://github.com/manisandro/gImageReader

[2] https://github.com/probonopd/gImageReader/releases/tag/conti...

Most people are best served by the big vendor OCRs. In my experience Amazon’s works the best, followed closely by Microsoft and Google at a distant third.

When did you do this comparison? A couple years ago I did a comparison and found Google the best and Amazon to be not very good.

Agree it’s best to skip tesseract unless the free cost is important. We spent a lot of time trying to preprocess and tune tesseract before realizing cloud OCR solutions are much better and fairly cheap.

I have been tasked with developing a Textract tool at work and so far I have been impressed with the complete accuracy of it for non-handwritten, non-photocopied documents.

I haven't seen it make any mistakes at all and responses take less than 3 seconds usually.

I tried it a couple of years ago on some Japanese receipts and it couldn't handle the mix of Japanese and English words/characters. Perhaps it was the way I set things up but the result was that of failure.

It seems the project has made a trade-off about language support.

One approach would be to say language doesn't matter, just train on converting any character from any language alphabet from image to text. The problem is that higher accuracy can be achieved by isolating characters from each language from each other. I imagine that particularly for Latin alphabet languages, accuracy must improve dramatically by splitting out any kanji or hanzi.

Even directly for Latin alphabet languages you would typically want to know the language, both to reduce the charset (i.e. the distinctions between smudgy ã or â or ā or ä or á often are trivial if you know the language but may be quite hard to tell otherwise) and to use a proper language model for disambiguating individual character guesses.

I also failed at using this. I think it needs training data and the default set leaves a lot to be desired. There are also hundreds of options which makes it difficult to wrap your head around. I was reading from screenshots, black English text on white background with a TrueType font. Never worked.

I hacked something together with "Capture2Text". Basically taking the screenshot, saving to jpg, shelling to the exe and getting the text back. Works pretty good.

Reply on news.ycombinator.com