How to improve accuracy for OCR?

[Peter]

Hello.

I'm trying to train Tesseract for OCR. My goal is to be able to recognize text from MRZ zone of various documents (mainly national ID). The training process should be pretty straightforward and I'd expect good results since all I have to deal with is one font (OCR-B), capital letters of Latin alphabet (A-Z), digits 0-9 and "less than" sign (<). Unfortunately the results are worse than expected. While the effects for preprocessed images (thresholding using GIMP) are pretty good (not perfect - in many cases Tesseract treats 0 as O, sometimes treats 5 as S and occasionally inserts unexpected whitespace between letters), data taken directly from an unprocessed scanned image is rather poorly recognized. There are many cases where Tesseract thinks O is 0, 5 is S, 8 is S, 2 is Z, H is M, U is W etc. While 0 vs O case can be tough (OCR-B 0 and O don't look too different) and perhaps beyond Tesseract capabilities, I think other ambiguities can and should be eliminated. As I'm new to Tesseract (have been using it for just a few days now) I hope you can suggest me the optimal training for OCR-B font or even provide me with some good training sample. Here's what I did to train Tesseract:

1) Prepared training text with OCR-B font (train1.odt, see attachments), converted it to .pdf with LibreOffice Writer (train1.pdf, see attachments)
2) Opened train1.pdf in GIMP and saved it as 300 dpi tif (resolution: 2479 x 3508), can't attach it as its size is more than 30 MB
3) Prepared font_properties file with the following line: ocrb 0 0 1 0 0
4) Executed the following Tesseract commands:

tesseract mrz.ocrb.exp0.tif mrz.ocrb.exp0 batch.nochop makebox (corrected mrz.ocrb.exp0.box included as attachment)
tesseract mrz.ocrb.exp0.tif mrz.ocrb.exp0 box.train
unicharset_extractor mrz.ocrb.exp0.box
shapeclustering -F font_properties -U unicharset mrz.ocrb.exp0.tr
mftraining -F font_properties -U unicharset mrz.ocrb.exp0.tr
cntraining mrz.ocrb.exp0.tr

changed shapetable, normproto, inttemp, pffmtable, unicharset names so that they're prefixed with mrz

combine_tessdata mrz.

To read data from images, I run:

tesseract -l mrz <image_file> output

Is there anything that can be done better? Maybe my training file is not good enough?
If that's the case, can you please provide me with a better one?

One more unrelated question. How to read data from image with non-standard orientation
(upside down, rotated left/right by 90 degrees)? How to use OSD feature?

[Nick White]

Hi Peter,

Sorry for the lack of response, I think us regulars here are all
quite busy at the moment.

Have you searched the archives of this mailing list? I seem to
recall someone previously deciding to go with a different project
which focused just on MRZ recognition.

Tesseract will do a reasonable job, as you have found, but perhaps
a dedicated program could do even better (and for less effort on
your part).

As far as improving your Tesseract results, though, I'd recommend
looking into user_patterns. It isn't well documented, but if the
format you're expecting is predictable it should help. Also have you
set up a unicharambigs file? That may help a little too (not much,
but it's probably worth adding for the common cases of 5 -> S, 8 ->
B, etc).

> One more unrelated question. How to read data from image with non-standard
> orientation
> (upside down, rotated left/right by 90 degrees)? How to use OSD feature?

I confess I don't actually know. I think Tesseract might try to
guess this entirely by itself. Does anyone else here know any
better?

Once you're happy your MRZ training is as good as it will get, would
you be happy to have it added to the main Tesseract repository? If
so (and it'd be great if you were) open an issue on the bug tracker
with the training file, and add some comments to the top of
mrz.config about how it was created and where the source files for
it are (see my grc.traineddata for an example).

Thanks Peter, and sorry again for not getting back to you sooner,

Nick

P.S. One other thing I just thought of: is the DPI you're feeding
into Tesseract the same as the DPI you trained with (300)? Ideally
it should be. Also you're right to preprocess using thresholding;
Tesseract isn't particularly good at that step and it's much better
if you can do it first.

> --
> --
> You received this message because you are subscribed to the Google
> Groups "tesseract-ocr" group.
> To post to this group, send email to tesser...@googlegroups.com
> To unsubscribe from this group, send email to
> tesseract-oc...@googlegroups.com
> For more options, visit this group at
> http://groups.google.com/group/tesseract-ocr?hl=en
>
> ---
> 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.
> For more options, visit https://groups.google.com/groups/opt_out.
>
>

[Zsombor Kaló]

To whom it may concern,

Just created a tesseract 3.04 trained data (see attached). 

I call it "MRZ" because it has the OCR-B as the only font and trained with characters A-Z, 0-9 and the lesser-than symbol (<). Seems to be fast and accurate in my projects.

[Jasnan Tp]

hi,

When I use mrz.traineddata, I get the following error

tesseract test.png result.txt -l mrz

Tesseract Open Source OCR Engine v3.03 with Leptonica

index >= 0 && index < size_used_:Error:Assert failed:in file ../ccutil/genericvector.h, line 589

[1]    15786 segmentation fault (core dumped)  tesseract test.png result.txt -l mrz

Is this because mrz.traineddata is corrupted?

[Quan Nguyen]

It was said created for 3.04. Can you try it with Tesseract 3.04?

[Wilko Meijer]

I had the same problem. Eventually I created a traineddata file from the OCR-B ttf, which works great for me. See attached.

[Jasnan Tp]

Wilko,

thank you very much for sharing your traineddata. it works almost perfect for me. most of the time 100% correct results and otherwise at least getting 98% right. May I know how you trained tesseract? the tools you were using to build this trained data?

Thanking You

Jasnan 

[Wilko Meijer]

By using this website: http://trainyourtesseract.com/

I tried creating traineddata the normal way with tesseract too. But since I didn't have many images to use for the traineddata and didn't have much time, I opted for this solution.