Preprocessing - detailed cropping

[Mirko P]

Hello,

I'm trying to recognize the machine readable part of a passport. (see the last line in this picture: http://s.hswstatic.com/gif/passport-11.jpg )

I'm using Tesseract on Android (tess-two) and take the picture with a 5 Mpix mobile camera. Unfortunately, the accuracy is not satisfyingly high. What I have tried to improve recognition was cropping the picture and retraining Tesseract for the font used in a passport (ocr-b). Both raises accuracy but still not to an acceptable level.
Here is a typical cropped picture I hand to Tesseract to perform ocr:

The binarized picture created by Tess for the actual recognition looks like this:

This is what Tesseract recognizes:

  09 1 M 1 907 1 8  F8 F857<4 < W<B<O <UME
  QVWBBENO W JMGHJ <RBP6W9BQR ED


I figured that the thin line at the bottom is extremely distracting to Tesseract. If I cut off the line manually and perform ocr, results are perfectly fine and all characters are recognized.
My question is, how can I find and get rid of that line automatically if it is in the cropped picture? This has to be done on an Android phone.

Any help will be appreciated!
Mirko

[Mirko P]

Anybody any idea? Is blob detection the way to go? Unfortunately I don't know too much about image processing and its possibilities.

[Dmitri Silaev]

Connected component (CC) analysis in the binary image is the way to go. See this for theoretical introduction: http://en.wikipedia.org/wiki/Connected-component_labeling . Then you have two approaches (can be used simultaneously):
  - Filter only good CCs, i.e. having size above minimum defined width and height, or

  - Discard bad CCs, i.e. having pixel count less than defined, or with longest principal axis wrongly directed or too lengthy (look for best-fitting ellipse, ellipse of inertia and so on), Y centroid's coordinate, etc.

In principle, there are tons of possible CC properties that could be used to filter or discard good or bad CCs, and their combinations. Much room for experimentation in you specific problem. I know Leptonica has a few routines to work with CCs. You can also try OpenCV (blobs in their terminology). I suppose, however, some code you would need to write yourself.

HTH

Best regards,
Dmitri Silaev
www.CustomOCR.com

On Fri, Apr 10, 2015 at 11:50 AM, Mirko P <mirko.p...@gmail.com> wrote:

Anybody any idea? Is blob detection the way to go? Unfortunately I don't know too much about image processing and its possibilities.

--
You received this message because you are subscribed to the Google Groups "tesseract-ocr" group.
To unsubscribe from this group and stop receiving emails from it, send an email to tesseract-oc...@googlegroups.com.
To post to this group, send email to tesser...@googlegroups.com.
Visit this group at http://groups.google.com/group/tesseract-ocr.
To view this discussion on the web visit https://groups.google.com/d/msgid/tesseract-ocr/e7c0bdf9-f546-438a-b536-b748968104c5%40googlegroups.com.

For more options, visit https://groups.google.com/d/optout.

[Mirko P]

Thanks a lot! I'll try that.