Testing The Accuracy Of Google Docs OCR

A potentially handy feature of Google Docs is its option to perform optical character recognition (OCR). As we've suggested in the past, that offers a convenient and free way to convert scanned images into text. But just how accurate is it?

I recently had to scan a dozen or so pages of introductions from various novels, which I wanted rendered as text. That represented a good opportunity to test just how well Google's OCR system actually worked. Scanning the images (at a high resolution) and uploading them was relatively quick, taking a total of perhaps 20 minutes.

However, despite the fact that I was working from a professionally-printed book, the results weren't spectacularly impressive. I ended up with around 2,896 words of text, but even a quick glance showed there were lots of text artefacts and misinterpretations. Here's a fairly representative sample paragraph:

One day, having lunch. at a C01"11.er Hnuse, ji was emaptured by a conversation un statistics going my at a tabie behind. me. I turned my he-ad auf? camght a vague giimpse of a bald head.,

Editing all the text to remove all the errors took me a further 30 minutes. Some mistakes were fairly mundane and easily fixed (Google tended to interpret the letter 'i' in my original documents as the number '1', sometimes added extra spaces between words, and tended to overenthusiatically add punctuation, as you can see in the example above.) Some misinterpretations would have been difficult to correct without the reference images, and on several occasions Google's OCR engine simply ignored entire words, a problem that would not always have been obvious without a careful comparison.

Given that the Docs OCR option is free, experimenting is not a costly exercise. However, with a total time of 50 minutes spent on getting the text scanned and accurate, it wasn't as much of a time saving over typing up the material myself as I had originally expected.

How useful have you found Google Docs' OCR capabilities? Share your own experience in the comments.


    Angus, Did you compare this on windows or other OCR with the same document?.

      No -- but it's a fair question and something I might look into down the road.

    Mark, I'm not really sure why running it on Windows would make any difference to the performance of a cloud based service - but I would like to see how the same document goes with other OCR software as you suggested.
    This is exactly the sort of thing I would have expected Google to excel at.

      Sorry, I mean the microsoft office document imaging OCR on windows

    My occasional experience with OCR's over many years has been distinctly underwhelming. There have been many instances where I would have been better offer retyping an albeit modestly sized document from scratch rather than scanning, OCR'ing and then laboriously error-correcting. Anyone out there willing to speak up for this technology?

    I would be interested to see this compared to the OCR in Evernote.

    Gus, I've been more than happy with the OCR performance of Evernote Premium. I've bought the premium subscription for several years now (initially just as "thank-you money), and have been really happy with its ability to not only OCR documents, but also words in photos, etc.

    A handy feature of this article would be a scanned copy of the document so we can see for ourselves how good the quality of the input text was.

    I found OCR software far more responsive when given greyscale scans instead of colour or black and white.

    You can always use services like www.kirtas.com or www.bookscanning.com . The output ( documents ) would be professionally compared against the original.


    Angus. I am a big OCR fan but I have used paid versions for many years now in order to achieve high quality. I read your report with interest, but your average of 60 words per minute is very slow compared with what I get out of my scans. I used OCR heavily in my higher education studies, and now use it for agreements and information gathering weekly. I also scan books for future search and reference. If you would like to send me your original image, the one you showed the result for above, I'll run some testing here.

    Alternatively, you can test yourself. WiseTREND offers a free trial of high quality Online OCR, which you can use over e-mail or API. See here:
    Send your same image to [email protected] and see what you receive back.

    Let me know in either way.

    If you're on Windows, you can send a screen capture from Snagit to the Microsoft Office Document Imaging (MODI) app for OCR.

    This video tutorial shows how to set up a custom profile to accomplish that: http://www.screencast.com/t/Waf5S8fnPh

    I might add that I've always had nearly perfect results with MODI - even when words were scattered about a pictorial chart image.

    It does require you to have Microsoft Office to access MODI... though, you can upload your image to Google Docs but the accuracy is just not there, yet.

Join the discussion!

Trending Stories Right Now