From #83, @zdenko's quote:

If you run:
tesseract OCR.tif ORIGINAL pdf
than ORIGINAL.tif is included in ORIGINAL.pdf WITHOUT any modification.

should be readed as: OCR.tif is included in ORIGINAL.pdf WITHOUT any modification. ;-)
And OCR is run on OCR.tif
IMO dhendrix what something else - he want to run OCR on image_b (improved for OCR), but include image_a (original) to pdf...

zdenop: Yes, that is correct. I want to run OCR on image_b (improved for OCR), but include image_a (original) in the resulting PDF.

BTW - I just realized that there is a user forum (https://groups.google.com/forum/#!forum/tesseract-ocr). Maybe somebody has asked / answered my question there. My apologies for not looking at that forum earlier.

Please post an example of a cleaned vs uncleaned image where accuracy improves significantly. Or even better, point to some documentation that has some examples. In the long term, one would hope that OCR could improve such that having a separate cleaned image is unnecessary.

Regarding this feature request, I think it is probably better to use an outside utility that can replace the images in a Tesseract produced PDF. The caller is already generating a separate set of clean images, so is therefore comfortable with pre/post processing. This approach lets us keep the design intent and implementation of PDF generation simple ('don't mess with the images'). I don't know if such a tool exists already, but based on my knowledge of the Tesseract PDF it shouldn't be too hard to write. Apologies, but I am not volunteering to write one unless I need it myself for something. The closest existing thing I know about is OverlayPDF from Apache PDFBox. Previously mentioned here. https://www.mail-archive.com/[email protected]/msg11853.html

Also, if you really want to hack Tesseract to do what you are asking for, the code is in api/pdfrenderer.cc. You would have to replace both the pix and the filename. I'd just be reluctant to make this a general feature of Tesseract.

https://github.com/tesseract-ocr/tesseract/blob/master/api/pdfrenderer.cpp#L894

Hi jbreiden,
There are lots of tutorials on-line for how to clean-up images for improved accuracy OCR out there, just use your favorite search engine. Here's a good one from 2014 (not too old): http://www.christophermchurch.com/my-struggles-with-ocr-and-microfilm-scans/ . Yes, it would be nice if OCR engines were perfect and didn't need cleaned-up images. But that's currently not a realistic expectation when papers get crumpled, have background images, have shaded regions, are "scanned" using a cell phone camera with poor lighting conditions, etc.

I don't think it would be wise to try to add all that clean-up functionality into tesseract, which is why I'm proposing a solution to take an image that has already been processed externally. The exact intent is to "not mess with the images."

An external tool that could replace the image layer would certainly be good, but I haven't found any (suggestions welcome!). I tried to use hocr2pdf to use tesseract's .hocr data from my cleaned image and add it to a PDF with my original image but ran into a showstopper issue - When searching for a word in the document, the PDF viewers I tried would highlight the wrong part of the document. Maybe there is a bug with hocr2pdf or I am using it incorrectly. Tesseract already knows how to make a PDF so this reduces the possibility of an external program interpreting the PDF or hOCR specs differently and ruining the output.

Anyway, thanks for the pointer to api/pdfrenderer.cc! I might just add the feature locally to address my needs, or maybe try to start a new program based off of it as you suggest.

I don't mind closing this issue if others feel this feature is inappropriate or a better solution is made.

Maybe this tool could help.
https://github.com/tmbdev/hocr-tools/blob/master/hocr-pdf

Swapping images in Tesseract a PDF is pretty easy for a programmer if destination images are JPEG or JPEG 2000. It is really just a matter of cutting and pasting the data, then cleaning up the results with qpdf. The hardest part is getting the courage to open up a PDF file and look inside it.

Regarding HOCR and bounding boxes, make sure you have image resolution metadata set correctly everywhere. The hocr-pdf program mentioned above works okay, but is limited to latin character sets and will also struggle with ligatures in English.

Heh, yeah, I opened up a PDF and saw the stream content for the image and was thinking of how to replace it, but it seemed like there'd be a non-trivial amount of work to get right for a PDF n00b such as myself.

As far as I could tell the image resolution metadata was correct, or at least consistent. Couldn't get hocr-pdf working unfortunately due to some python module dependency that I couldn't find (I installed PyXML but no dice).

In the end I just went ahead and hacked my feature into Tesseract and it works pretty well :-) Here it is if you're interested, though be warned that it's a bit of a kludge in its current state: dhendrix@6cc206f

Thanks for the helpful pointers! Feel free to close if this feature is not desired for upstream, it can live on in my github account.

The python dependency is reportlab.