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. On Wed, Jun 19, 2013 at 11:45:10PM -0700, Peter wrote: > 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? > > > > -- > -- > You received this message because you are subscribed to the Google > Groups "tesseract-ocr" group. > To post to this group, send email to [email protected] > To unsubscribe from this group, send email to > [email protected] > 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 [email protected]. > For more options, visit https://groups.google.com/groups/opt_out. > > -- -- You received this message because you are subscribed to the Google Groups "tesseract-ocr" group. To post to this group, send email to [email protected] To unsubscribe from this group, send email to [email protected] 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 [email protected]. For more options, visit https://groups.google.com/groups/opt_out.
