Hacker News Reader

HNHacker News

Top New Best Ask Show Jobs

Comment by Synergyse | Hacker News Reader

Full thread

Synergyse·

Nice idea, would it be possible to do some OCR at the end?

View on HN

I did experiment with using this to bring out the text in photos of books. Here's an example:

Input image - http://i.imgur.com/6o5FwxG.jpg

Output image - http://i.imgur.com/7OIOxfO.png

I tried to use vanilla Tesseract on it, but I had no luck getting anything usable out of it.

You have to compensate for the 3D deformation present in the captured image. There has been some impressive work in this area recently.

Some commercial OCR engines such as Nuance Capture SDK have built in functionality for this.

Take a look at Tesseract, which last time I looked was the best codebase to use in this area. It's part of Google's open source multilingual OCR suite, which is in two parts (layout analysis and actual OCR code being segregated): https://code.google.com/p/tesseract-ocr/

Is there a guide to using tesseract? When I last used it, I had trouble accurately recovering text from an image I just created with Times New Roman. There are probably some settings I'd need to change to get it to work properly.

Not sure. I looked at it years ago, during its early open source stages, and similarly concluded that it was nontrivial to get going. Should be easier now. IIRC in those days it had a very high volume mailing list for user support.

At the end the script needs to upload the image to mechanical turk for analysis.

Reply on news.ycombinator.com