getting started, but no results
94 views
Skip to first unread message
michael c
Dec 28, 2021, 4:28:12 AM12/28/21
to tesseract-ocr
I am just starting to use the tesseract package and having no luck getting it to recognize anything.
My environment is C# using the package from nuget.
I am able to run fine, just no text is recognized in my sample image.
It does work on the provided 'phototest.tif'.
I have fiddled with many parameters in the config file and none has resulted in any useful output.
I only need to recognize digits and the image will have the same consistent form as this one.
Any hints on parameters i should look at to get this running?
My environment is C# using the package from nuget.
I am able to run fine, just no text is recognized in my sample image.
It does work on the provided 'phototest.tif'.
I have fiddled with many parameters in the config file and none has resulted in any useful output.
I only need to recognize digits and the image will have the same consistent form as this one.
Any hints on parameters i should look at to get this running?
Zdenko Podobny
Dec 28, 2021, 4:33:56 AM12/28/21
to tesser...@googlegroups.com
Did you read the docs?
ut 28. 12. 2021 o 10:28 michael c <thisism...@gmail.com> napísal(a):
--
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/d12cad94-ad9c-4659-87bc-94a57b58a4e1n%40googlegroups.com.
michael c
Dec 28, 2021, 3:45:54 PM12/28/21
to tesseract-ocr
Thanks, yes i had looked at that.
I began by expanding the image by 5x to get the characters to about 50 pixels high (vs about 8 initially).
My initial tests generated a tessinput.tif that looked very good to my eye, but did not work for the OCR.
I ended up also doing:
- posterize to level 2 to reduce the colors
- dilate, to reduce the thickness of the characters
But this still was not working.
I suspect the single characters and the lines between are causing issue.
I had tried several of the many settings in the config file hoping to figure out ones that would work but got nowhere and seemed to be shooting in the dark.
As i am unfamiliar with these many settings and did not find details on the meaning of many of them, my question was hoping to find some ideas on which ones might be helpful.
In the end i have defined rectangles for the position of each character, then copy all these rectangles to a new image placing characters in nice rows.
This worked on the sample image i have but i do not yet have additional samples to see if it works on them.
I had hoped to avoid coding the detail for the exact position of the characters and that it might read them as is.
Will see later when more samples arrive if this is a workable solution.
I began by expanding the image by 5x to get the characters to about 50 pixels high (vs about 8 initially).
My initial tests generated a tessinput.tif that looked very good to my eye, but did not work for the OCR.
I ended up also doing:
- posterize to level 2 to reduce the colors
- dilate, to reduce the thickness of the characters
But this still was not working.
I suspect the single characters and the lines between are causing issue.
I had tried several of the many settings in the config file hoping to figure out ones that would work but got nowhere and seemed to be shooting in the dark.
As i am unfamiliar with these many settings and did not find details on the meaning of many of them, my question was hoping to find some ideas on which ones might be helpful.
In the end i have defined rectangles for the position of each character, then copy all these rectangles to a new image placing characters in nice rows.
This worked on the sample image i have but i do not yet have additional samples to see if it works on them.
I had hoped to avoid coding the detail for the exact position of the characters and that it might read them as is.
Will see later when more samples arrive if this is a workable solution.
Yesbird
Dec 28, 2021, 5:23:46 PM12/28/21
to tesseract-ocr
I beleave that key is page segmentation mode - try to play with it:
https://www.pyimagesearch.com/2021/11/15/tesseract-page-segmentation-modes-psms-explained-how-to-improve-your-ocr-accuracy/
https://www.pyimagesearch.com/2021/11/15/tesseract-page-segmentation-modes-psms-explained-how-to-improve-your-ocr-accuracy/
Zdenko Podobny
Dec 29, 2021, 2:42:09 PM12/29/21
to tesser...@googlegroups.com
No. The problem is document layout and text detection.
Tesseract is OCR engine - it can detect text but on images with simple layouts (books page, newspaper, etc).
The input image is too complex, so for a good result you need to send to tesseract API only text region.
If the input image is the same with the same text position (only numbers are changed), it should be easy to define text regions in advance.
If text position could change, the solution would be to remove all backgrounds colors, graphics, lines. Anyway, the tesseract will look for lines, words etc, so the output could be not satisfying.
Another solution would be to use external text detection (e.g. YOLO see https://medium.com/saarthi-ai/how-to-build-your-own-ocr-a5bb91b622ba)
If you search only for numbers with the same font, you can consider creating a simple OCR with opencv (https://stackoverflow.com/questions/9413216/simple-digit-recognition-ocr-in-opencv-python, https://github.com/goncalopp/simple-ocr-opencv - example is in python, but you can get idea behind)
Zdenko
ut 28. 12. 2021 o 23:23 Yesbird <syane...@gmail.com> napísal(a):
To view this discussion on the web visit https://groups.google.com/d/msgid/tesseract-ocr/a80e3318-f8a6-4c4e-b6d7-4401174825b1n%40googlegroups.com.
Reply all
Reply to author
Forward
0 new messages