September 29, 2009

Google Docs OCR

Google Docs API tests a new feature that lets you perform OCR (optical character recognition) on an image. There's a live demo that illustrates this feature: you can upload a high-resolution JPG, GIF, or PNG image that has less than 10 MB and Google Docs extracts the text and converts it into a new document. Google mentions that "the operation can currently take up to 40 seconds" and a small test showed that the service is not yet reliable: it's slow and it frequently returns errors.


The results are far from perfect and you'll find many errors, but the service is free and it's constantly improving. Here's the result of the OCR for this scanned document:


There aren't many free OCR services available, so an OCR service provided by Google would be very popular. ABBYY FineReader Online is one of the best online OCR services, but the free version is limited to 10 pages a day.

Google sponsors the development of an open-source OCR software called OCRopus, but it's not clear if the online service provided by Google Docs uses OCRopus.

25 comments:

  1. This fits in very will with their acquisition of reCAPTCHA....

    ReplyDelete
  2. @ebboodah — yup, then can send hard words to users od reCaptcha to get amazing accuracy. And soon after that, cybercriminals will invent API to send reCaptcha images to Google Docs OCR for great accuracy of spamming. Perpetuum mobile. :D

    ReplyDelete
  3. Great idea ! But you're right : this doesn't work very well. I 'm excpecting a much more performant version :).

    ReplyDelete
  4. Is this English only for now? Currently we're very pleased with the Acrobat tools performance, but that installation is not available to all users. Web based would be great.

    ReplyDelete
  5. If you are going to criticize Google for not having perfect OCR translations, at least get the spelling right in your post.....

    "excpecting a much more performant version :)"

    ReplyDelete
  6. Is this like stegnography?

    ReplyDelete
  7. I just tested it on some pictures. In some cases, I need to resize *2 the image, to be recognized, but it's awesome to be able to use OCR online !
    Don't forget OCR is very complex technology, so, be patient !

    ReplyDelete
  8. Very poor quality

    The best online ocr service is :
    http://www.finereaderonline.com

    ReplyDelete
  9. Google and OCR technology. My prediction is pretty close:
    http://detailcode.com/news/future-predictions-of-googles-recaptcha/

    Has anyone tested the accuracy when images are slightly skewed/rotated?

    ReplyDelete
  10. Dear Google,

    it would be so kind of you if you could further in your endeavour and make Google Docs OCR work for Hebrew and Arabic. Perhaps, Cyrillic too. Certainly, people who read Chinese, Japanese or Hindi have similar levels of enthusiasm too.

    This is great news, as I have a lot of books I wish to scan. Contrary to people who love collecting books on shelves but not reading them, I love reading books but not keeping them. Printed books are real heavy stuffs and inconvenient to move around with. Printed books also do not lend themselves willingly towards searching their contents - very bad feature.

    I want to be able to read my stuffs anywhere I am in the world. I would scan the books, recycle the paper, but keep the covers in a trunk in an obscure corner of the cellar as proof of my ownership of the books.

    BTW, what's the meaning of "performant"? I am afraid the frequent appearance of this word is about to make it join the list of annoying words like "obligated" that have replaced "obliged", "percentage points" instead of "percent" or "presumptive candidate" instead of "presumed candidate" as real words.

    ReplyDelete
    Replies
    1. Actually, "percentage points" is an important concept, with a different meaning from "percents". For example, let's say that in one year, unemployment was 10%, and in the next year it rose to 11%. You might say that "Unemployment rose by 1 percentage point", or that "Unemployment rose by 10 percent" (because 1% is 10% of 10%; the number of unemployed rose by 10%). It would be confusing and *wrong* to say that "Unemployment rose by 1 percent". This is why the term "percentage point" is needed.

      Delete
  11. Jurgi,
    That way, Google can find the criminals much faster and black-list all their property faster.

    Hence,
    Isn't “candidate” enough?

    ReplyDelete
  12. Good start, the other so called free services suck.

    ReplyDelete
  13. Tesseract seems to have some quality and formatting issues.

    ReplyDelete
  14. why is this redirecting me to googlecodesamples.com which doesn't appear to be owned by google? And then asks me for access to my docs. Is this legit?

    Registered through: GoDaddy.com, Inc. (http://www.godaddy.com)
    Domain Name: GOOGLECODESAMPLES.COM
    Created on: 30-Jun-08
    Expires on: 30-Jun-10
    Last Updated on: 01-Jul-09

    Private, Registration GOOGLECODESAMPLES.COM@domainsbyproxy.com
    Domains by Proxy, Inc.
    DomainsByProxy.com
    15111 N. Hayden Rd., Ste 160, PMB 353
    Scottsdale, Arizona 85260
    United States
    (480) 624-2599 Fax -- (480) 624-2598

    Domain servers in listed order:
    NS1.SLICEHOST.NET
    NS2.SLICEHOST.NET

    ReplyDelete
  15. Tesseract has a number of formatting and accuracy issues. So far, the best solution I have encountered is OCR Terminal (www.ocrterminal.com). 20 free pages a month and very good accuracy.

    ReplyDelete
  16. What's the point of this if you can't upload a PDF file.

    ReplyDelete
  17. Try this online OCR instead. Worked great for me.
    http://www.cvisiontech.com/index.php?option=com_docman&task=cat_view&gid=45&&Itemid=206&lang=eng

    ReplyDelete
  18. I think that they are using neither Ocropus nor Tesseract. When i gave the same image to Google OCR, Tesseract, and Ocropus, I noticed that Google OCR provided better results than Tesseract, and Tesseract provided much better results than Ocropus, although Ocropus uses Tesseract as the recognition engine....strange!

    I nearly tried all free online OCR services. All of them give more or less the same quality. I really like i2OCR because it provides a balance between accuracy and usability.
    http://www.sciweavers.org/free-online-ocr

    Finally, the secret sauce is DPI. The more DPI, the better the quality of recognition.

    ReplyDelete
  19. can u pls let me know how can i use api code for my own stuffs to read text from image

    ReplyDelete
  20. HELLO YUVATANTRA,
    i also have the same pb, i need to use the api code for my ownn project.
    Are u a developper?

    ReplyDelete
  21. Well, I just tried Google's OCR and it was on pants at b/w 600dpi. Junk character"5 and no formatting. This may have been down to the font (bold Tahoma, don't ask me why, some people just have no sense of style). However, tried the same document with onlineocr.net, and got much better results.

    ReplyDelete
  22. Please make this service available in Arabic... good start anyway.

    ReplyDelete
  23. You should try http://www.ocr-extract.com/

    ReplyDelete

Note: Only a member of this blog may post a comment.