Ask HN: What is the best open source OCR software supporting multiple languages? | Hacker News Reader

Ask HN: What is the best open source OCR software supporting multiple languages?

86 pointspostila38 comments

Tesseract[0] is a system that is broken in to different parts, at least one does layout analysis and another does the actual OCR. Output is a different layer again. I believe it is an open source adaptation of what Google used for its books project. The interface was less than polished a few years ago, to the point where getting it running at all was rather difficult. However, for multilingual work (including Chinese) it is probably ideal.[1] Note that if you are scanning books there are now some interesting open hardware systems appearing online that turn pages and take photos with cameras, so you can scan books - without cutting them up - to a high resolution.

[0] https://github.com/tesseract-ocr/tesseract [1] https://github.com/tesseract-ocr/langdata

My master thesis[1] was about OCRing a multilingual dictionary (Vietnamese-French using Nôm characters) and I got the following results, using built-in language model:

- OCR for French is very good

- OCR for Vietnamese (latin characters) was dispointing

- OCR for Chinese indeed outout Chinese characters but not the good ones

- OCR for Chinese didn't handle letters in Chinese script

- no support for Vietnamese Nôm (not suprising)

The good news is: - you can train new language model

- the software is built on a (clean?) API you can use yourself

- there is some usable documentaition to get started & also for training models

- headers files are well commented

So this enable one to build a solution on Tesseract, but this is by no way an out-of-the-box solution especially if you have very specific needs. All in one it is the best free (libre, 0$) OCR engine out there.

[1] http://crim.fr/sites/default/files/memoire_m2_Lecailliez_fin... [French]

Tesseract can give you nightmares, but unfortunately it's the only solid OCR library out there.

What kind of nightmares?

Beside Tesseract which was a state-of-the-art OCR software by HP in the early nineties and recovered by Google a few years ago and is open source.

There is Cuneiform, a former main competitor to ABBYY Finereader. CuneiForm got open sourced a view years ago, though in a sad state (project files where in VS C++ 6 ('98), comments in Russian), but a community fixed that and ported it to Linux. It's also probably the best one for Russian language. It also has an UI and some advanced features that only ABBYY amd Cuneiform have, but non of the other competitors (certainly no other open spurce OCR package). https://en.wikipedia.org/wiki/CuneiForm_(software)

Sorry to hijack, but what about the best OCR service? I'd much rather farm the OCR work out to another service than trying to do it myself.

I bought ABBYY FineReader for Mac for abut $99. I find it to be pretty amazing. My new scanner also came with it and I generally expect it to do a reasonable text translation of whatever I throw it, whether it be newsprint articles or crumpled receipts.

If you need to do OCR that also preserves table structure -- which is what I bought FineReader for in the first place, I don't think there's any open source alternative, and FineReader does a very capable job.

Here's an example of FineReader in action: OCRing the docs released by the FBI on Clinton's email system. I've also included the pdftotext output showing how FineReader's text conversion also attempts to preserve the physical layout of the text characters:

https://github.com/dannguyen/clinton-hillary-email-fbi-inves...

They also have a Linux CLI for servers, it seems to be licensed by annual page volume, https://www.abbyy.com/ocr-sdk-linux & http://www.ocr4linux.com/en:pricing

thank you, I'll check it out. Know this company very well, we originated from the same university, but I've completely forgotten to check their new products now. Thanks.

Oh. For 500.000 pages per year the ask >5k euro... No way. I need to process ~50M images right now, and then ~50k every day.

Make them an offer. I can almost guarantee there's room for negotiation.

Try OCR.space - it is 10$ per 100,000 (and probably even less for your volume)

That's pretty good! How does it do with your handwriting?

I used to use Evernote quite heavily and I've never been able to replace their excellent OCR. I would search for some text and was always blown away when it would find a photo of a whiteboard or a sketch of mine.

Any idea what Evernote uses?

I'm not going to embarrass myself by showing my handwriting. So here's how it does on 3 famous handwritten letters (not well) https://gist.github.com/dannguyen/50dcb5a8f4230e7a8a40bfe2d0...

Since both ABBYY and Evernote were founded by Russians, I bet Evernote uses FineReader.

It is not the case. Evernote was created by https://en.wikipedia.org/wiki/Stepan_Pachikov who also founded Parascript company http://www.parascript.com, which was doing OCR long before Abbyy and is mostly specialized on handwriting. So both technologies are from Russia, but teams are independent.

Good question..

Thanks for the example. That's probably the best advertisement. I have a TON of physical notes that I want to turn into text. I know people use Evernote but I never trusted them. I prefer to keep it in-house and this is a solid solution. THANKS!

I use ABBYY for Linux: http://www.ocr4linux.com/

Everything about it sucks, cost, anual page limit license, I can only run it as root for some reason, but the OCR quality is superb.

here's the scripts I use if it's useful: https://github.com/cove/scanbd

Google Cloud Platform has an API for it: https://cloud.google.com/vision/

I have to say, this was the most accurate and quickest method I had tried when it came to OCR.

Here's an example (with Python code) of how it performs on street sign data:

https://gist.github.com/dannguyen/a0b69c84ebc00c54c94d

Unfortunately, it doesn't do well on large chunks of text if you're trying to OCR a scanned document.

Same result here. Google OCR shines on short text inside images such a streets signs.It works also well on mangas :-)

The Ocr.space OCR Api works well and has a very generous free tier. And then there is also the Google Vision API, which includes good but basic OCR (but does not support coordinates, just returns plain text)

I just tried two pictures with VINs in them. Both of them were successfully parsed by Google Vision, but OCR.space just errored out. It was not an encouraging result.

yeah.. Google OCR is better for more challenging tasks, but in my case (OCR on Screenshots and PDF OCR in Chinese and English) the ocr.space soluton works just fine, so I picked it as the most economical solution.

I've used Tessarect with Tess4J Java wrappers, which has been pretty good.

Not that good for Russian. For English, it's not the best as well – too many mistakes for some fonts.

Sorry, thought you meant multiple programming language support. Yes, definitely ran into some font issues and noise turning '1' into 'L' or 'T', etc. As people have been saying though, it may not be great out of the box for you but you can train it on the font you want.

Do you want to OCR several human languages, or do you want bindings/libraries in several programming languages? The question as written is a little ambiguous.

I need to process millions of images and extract texts from them as better as possible. Primary language is Russian, but some texts are English. Also, interested in other languages (Spanish, German, etc) for future needs.

What I've tried so far (including Tesseract) is either bad for Russian texts or cannot work with mixed texts (e.g.Russian with some English words). Or both.

Programming languages/platform don't matter, but smth Linux-compatible is better of course.

Tesseract can (and has to be) trained, so it can effectively support anything.

OCR isn't limited to language usually unless you are doing some really high end stuff when it does linguistic prediction but you only need that if you are working with really poor (image) quality sources.

But overall OCR is "language" agnostic, it is however usually not type set agnostic so what you would want to do is train it for whatever fonts are common for a particular language.

This gets slightly tricky if you have to do handwritten transcription or very stylized fonts but in those cases the "language" again is not an issue because your OCR program doesn't understand language to begin with.

Caffee Deep learning possibly outperforms Tesseract.

https://christopher5106.github.io/computer/vision/2015/09/14...

Tesseract

The best tool would be something that I can iteratively improve using some ML methods, that I would run on Linux and integrate into my programs. And open source, of course. I know, I want too much :)

No, not too much, at least I'd agree with you.

A lot of information about current Tesseract is there - https://github.com/tesseract-ocr/docs/tree/master/das_tutori... .

Tesseract is trainable, even though the bulk of capabilities came from algorithms designed well before deep learning became popular.

One of slides mentions that it's puzzling that Tesseract is "winning" over modern ML attempts to solve OCR. However, latest developments - adding LSTM networks to Tesseract - are reported to be promising. Wonder when they'll become available on the Github...

This looks interesting, thank you

Is there anything great (even if potentially pricey) for ICR (individual handwritten characters, usually separated by boxes) or handwriting? Preferably as a service.

Reply on news.ycombinator.com