Recognising numbers in sudoku

[agentbond009]

hello. I am working on a project, i was wondering if tesseract is an OCR can it read the no.s from a sudoku.
thanks

[Jean-Marc Spaggiari]

Hi,

Did you give it a try?

Try to have the numbers in white in black in a withe background. Are you able to get each cell separately?

JM

Le ven. 30 oct. 2020 à 10:22, agentbond009 <kshit...@gmail.com> a écrit :

hello. I am working on a project, i was wondering if tesseract is an OCR can it read the no.s from a sudoku.
thanks

--
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 view this discussion on the web visit https://groups.google.com/d/msgid/tesseract-ocr/4abe9a73-9543-4a5b-b29e-a917874378f8n%40googlegroups.com.

[agentbond009]

sorry for the late reply
I have done the thresholding part by using opencv and converted the whole image to gray
and still stuck with recognising the digits. i was thinking to use clear_border from skimage.

[Jean-Marc Spaggiari]

Is the image always the same size? Can you cut the columns to remove them and keep just the numbers?

To view this discussion on the web visit https://groups.google.com/d/msgid/tesseract-ocr/f2c178b6-5861-49be-8b23-3506f58c14edn%40googlegroups.com.

[agentbond009]

yes  i think the image is always the same size but i am not able to cut out the columns(still figuring out ). Like the image below is the result after thresholding. The problem is to reduce the noise(all depends on quality) like for example if the pic is of bad quality like this one, then i think the ocr will mess it up.
there are two pics for comparision.

[Martin Läuter]

Hi,

  Someone did some (apparently) solid sudoku solving without tesseract, see https://blog.scottlogic.com/2020/01/03/webassembly-sudoku-solver.html. And the explanation is not bad either.

  Hope that helps,

Martin.

[agentbond009]

Thanks Martin Läuter  for the tutorial. But still i am gonna use tesseract for this bc i don't think my PC will handle tenserflow.

[Jean-Marc Spaggiari]

If the pictures are all the same size, using ImageMagick you can most probably just cut each cell, pad it with black and send it back to tesseract. Just make sure to keep value only when there is a high confidence or enough "white" pixels?

To view this discussion on the web visit https://groups.google.com/d/msgid/tesseract-ocr/c66465e3-3640-4daa-9ae4-52ae2dbc8451n%40googlegroups.com.

[Lorenzo Bolzani]

Use hough lines detector to detect the lines and draw a thick white line over them.

https://opencv-python-tutroals.readthedocs.io/en/latest/py_tutorials/py_imgproc/py_houghlines/py_houghlines.html

Or use MSER to detect blobs of the appropriate size.

https://answers.opencv.org/question/19015/how-to-use-mser-in-python/

https://stackoverflow.com/questions/40078625/opencv-mser-detect-text-areas-python

Lorenzo

To view this discussion on the web visit https://groups.google.com/d/msgid/tesseract-ocr/c66465e3-3640-4daa-9ae4-52ae2dbc8451n%40googlegroups.com.

[agentbond009]

Jean-Marc i have posted the pic.

https://groups.google.com/g/tesseract-ocr/c/1pi2LR4ug_A/m/NM74kUO7DgAJ

[Martin Läuter]

Remember, it's a website, that is live-solving the sudoku! Tensorflow.js is pretty lightweight, and you only need it to evaluate the number recognition model. You don't need to train the model (the trained weights are checked in the repository, AFAICT) nor install TensorFlow at all.

  I don't say you should absolutely use that webassembly project, but there is nothing to be afraid there, the author did a pretty good job of preparing and documenting everything.

  Greetings,

Martin.