3

From your experience, what is the most accurate open-source Optical Character Recognition (OCR) library/software to read Japanese text?

I just tried nhocr, its mistake rate is over 2% even on an extremely clean high-definition document (2% is for ultra-clean characters in big font, for scanned books it is much worse, let alone handwritten forms).

Old Japanese cellphones (in particular Sharp) used to have great Japanese OCR 10 years ago already, so I guess it is feasible in Open Source too.

  • Is nhocr being actively developed? Also, have you looked at tesseract-ocr? Looks like it is a Google project now too. – user3169 Feb 4 '15 at 17:22
  • What about VietOCR? – Bob Ortiz Jun 21 '16 at 23:00
  • @EvanderConsus: VietOCR is a GUI frontend for Tesseract OCR engine, please anyone feel free to add an answer demonstrating how well Tesseract performs for Japanese :-) – Nicolas Raoul Jun 22 '16 at 2:43
  • Cellphones had OCR? Wow, I didn't know that. :) – Andrea Lazzarotto Sep 3 '16 at 19:36
2

You can use these suggested parameters to increase the accuracy of tesseract.

example usage from my jtesseract script:

tesseract "$@" -l jpn -c chop_enable=T -c use_new_state_cost=F -c segment_segcost_rating=F -c enable_new_segsearch=0 -c language_model_ngram_on=0 -c textord_force_make_prop_words=F -c edges_max_children_per_outline=40
1

I don't know any good open source software for this, but if that is not a hard requirement, I have used ABBYY FineReader 9.0 Express Edition for a couple of years now, and I am generally pleased with the quality of the OCR results. Not sure if I can say better than 98% but good enough to be easily correctable.
I use it mostly to OCR Japanese text from books or magazines. According to its recommendation, and my trial and error, very high resolution actually gives worse results than sticking to 300dpi.

The only consistant issues I came across:

1) Sometimes it will return a Chinese kanji that is similar to the actual Japanese kanji. I don't know if this can be restricted or just because the OCR supports both languages.

2) Some character might be incorrectly interpreted through an entire document. But after the first ones it is easy to Find & Replace.

3) It doesn't seem to be able to read English when mixed on a page. I just enter that when proofing. It looks to be up to v12 now. Possibly you could find an previous version for a good price.

0

From my experience I can recommend you two software which happen to be the best ones in the field of OCR but they are not Open Source.

  1. ABBYY

    Japanese OCR was first introduced by ABBYY FineReader. All versions of FineReader include support for Chinese, Japanese, Korean and Thai characters. ABBYY, a leading provider of document recognition, data capture and linguistic software, today announced the newest release of its FineReader 9.0 optical character recognition (OCR) and document conversion application with added support for Chinese (both traditional and simplified), Japanese, Thai and Hebrew languages.

  2. Readiris

    The latest version of ReadIRIS includes Japanese, Traditional Chinese, Simplified Chinese and Korean character recognition in its base packages.

I use this product for Asian documents because it almost supports all Asian languages. It's a Cannon company software and again it's not open-source.

I'm sure you'll get your desired results by using any of those two software.

  • 1
    Open source is a hard requirement. – Nicolas Raoul Feb 5 '15 at 2:57
  • do you want to modify the code? – Conspiria Feb 5 '15 at 17:21
  • Yes. All software has bugs and shortcomings. It would be much too risky to go with software that can only be modified by a single company that might choose to charge arbitrary prices for modification, or even refuse to modify at all. – Nicolas Raoul Feb 6 '15 at 2:03
  • that's right but what modification could you possibly make in a complicated software like OCR which is based on a library for supported languages? are you going to implement new compatibility to support your type of Japanese papers? if that's the case why don't you modify the 'nhocr' to remove the 2% mistake rate. – Conspiria Feb 6 '15 at 9:13
  • 1
    1) Yes, I am going to modify it, or pay a large amount of some company's money to another company to do it 2) Welcome to Software Recommendations, the place where people keep looking for better solutions :-) – Nicolas Raoul Feb 6 '15 at 12:01

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.