How to fix OCR errors?

[endolith]

This works pretty well, but is not perfect. How can I modify the invisible text to fix OCR errors without otherwise affecting the layout of the document. For example I want to change society. ' into society.¹.

I found a lot of people asking for this:

• [Feature]: Manually correcting OCR errors #1158
• ocr corrections #66
• Correcting recognition errors - possible with sidecar option? #781
• hocr import / export #453

[endolith]

One comment says:

With ocrmypdf, you could run --output-type hocr -k to generate the .hocr file which contains a HTML-like description of the recognized characters and edit this file then manually use hocrtranform to finish the file. It won't be exactly right unless you also fiddle with the pixel coordinates.

[endolith]

But that doesn't work:

ocrmypdf: error: argument --output-type: invalid choice: 'hocr' (choose from 'pdfa', 'pdf', 'pdfa-1', 'pdfa-2', 'pdfa-3')

[endolith]

Oh I see:

> ocrmypdf -l eng -k --pdf-renderer hocr "png2pdf.pdf" "ocred.pdf"
… 
Temporary working files retained at:
/tmp/com.github.ocrmypdf.7791wzq8

Which creates a bunch of files in the temp folder:

/tmp/com.github.ocrmypdf.7791wzq8:
000001_ocr.png        000003_ocr_hocr.pdf   000005_rasterize.png  000008_ocr_hocr.hocr  000010_ocr_hocr.txt   000013_ocr.png        000015_ocr_hocr.pdf   000017_rasterize.png  000020_ocr_hocr.hocr  origin.pdf@
000001_ocr_hocr.hocr  000003_ocr_hocr.txt   000006_ocr.png        000008_ocr_hocr.pdf   000010_rasterize.png  000013_ocr_hocr.hocr  000015_ocr_hocr.txt   000018_ocr.png        000020_ocr_hocr.pdf   pdfa.pdf
000001_ocr_hocr.pdf   000003_rasterize.png  000006_ocr_hocr.hocr  000008_ocr_hocr.txt   000011_ocr.png        000013_ocr_hocr.pdf   000015_rasterize.png  000018_ocr_hocr.hocr  000020_ocr_hocr.txt   pdfa.ps
000001_ocr_hocr.txt   000004_ocr.png        000006_ocr_hocr.pdf   000008_rasterize.png  000011_ocr_hocr.hocr  000013_ocr_hocr.txt   000016_ocr.png        000018_ocr_hocr.pdf   000020_rasterize.png
000001_rasterize.png  000004_ocr_hocr.hocr  000006_ocr_hocr.txt   000009_ocr.png        000011_ocr_hocr.pdf   000013_rasterize.png  000016_ocr_hocr.hocr  000018_ocr_hocr.txt   debug.log
000002_ocr.png        000004_ocr_hocr.pdf   000006_rasterize.png  000009_ocr_hocr.hocr  000011_ocr_hocr.txt   000014_ocr.png        000016_ocr_hocr.pdf   000018_rasterize.png  fix_docinfo.pdf@
000002_ocr_hocr.hocr  000004_ocr_hocr.txt   000007_ocr.png        000009_ocr_hocr.pdf   000011_rasterize.png  000014_ocr_hocr.hocr  000016_ocr_hocr.txt   000019_ocr.png        graft_layers.pdf
000002_ocr_hocr.pdf   000004_rasterize.png  000007_ocr_hocr.hocr  000009_ocr_hocr.txt   000012_ocr.png        000014_ocr_hocr.pdf   000016_rasterize.png  000019_ocr_hocr.hocr  images/
000002_ocr_hocr.txt   000005_ocr.png        000007_ocr_hocr.pdf   000009_rasterize.png  000012_ocr_hocr.hocr  000014_ocr_hocr.txt   000017_ocr.png        000019_ocr_hocr.pdf   metafix.pdf
000002_rasterize.png  000005_ocr_hocr.hocr  000007_ocr_hocr.txt   000010_ocr.png        000012_ocr_hocr.pdf   000014_rasterize.png  000017_ocr_hocr.hocr  000019_ocr_hocr.txt   optimize.opt.pdf
000003_ocr.png        000005_ocr_hocr.pdf   000007_rasterize.png  000010_ocr_hocr.hocr  000012_ocr_hocr.txt   000015_ocr.png        000017_ocr_hocr.pdf   000019_rasterize.png  optimize.pdf
000003_ocr_hocr.hocr  000005_ocr_hocr.txt   000008_ocr.png        000010_ocr_hocr.pdf   000012_rasterize.png  000015_ocr_hocr.hocr  000017_ocr_hocr.txt   000020_ocr.png        origin@

So then I have to modify the …_ocr_hocr.hocr files and not the …_ocr_hocr.txt files?

[endolith]

OK I made 72 manual small edits to the .hocr files… 😩 Now I need to figure out how to convert them back into a PDF.

[jbarlow83]

You can use the ocrmypdf API - api._pdf_to_hocr and then api._hocr_to_ocr_pdf to create the hocr files and then finish the PDF. It splits the normal pipeline into two steps. There is no command line option for this, because I don't think it would be very useful and I don't want to support people who try to manually edit hocr files - there are lots of ways you can mess up.

You could do small edits this way, but it mostly amounts to changing words in place. It would take a full word processor to do more complex changes. The APIs are currently private because they are subject to change.

The author of gscan2pdf is rewriting that program in Python and planning to use ocrmypdf's features, including an interface to edit hOCR files. No idea of his timeline.

[endolith]

Is that not available in the version installed by apt install ocrmypdf in WSL?

In [4]: ocrmypdf.__version__
Out[4]: '9.6.0+dfsg'

[endolith]

I tried the API but it doesn't seem to exist.

I tried the pdfbook solution in some comment but it won't build.

OMG why is it soooo difficult to just intercept the OCR and replace a few characters before it writes to PDF? This is like a day-long project just to replace a few characters.

[endolith]

Also not in the version installed by pip?

In [2]: ocrmypdf.__version__
Out[2]: '14.4.0'

[endolith]

OK I finally got it working on a different machine:

sudo apt install ocrmypdf
conda create --name ocrmypdf pip ipython
conda activate ocrmypdf
pip install git+https://github.com/ocrmypdf/OCRmyPDF.git
ipython

Then inside ipython:

import ocrmypdf
from pathlib import Path
ocrmypdf.api._pdf_to_hocr(input_pdf=Path("png2pdf.pdf"), output_folder=Path("./output"))
ocrmypdf.api._hocr_to_ocr_pdf(work_folder=Path("./output/"), output_file=Path("OCRed.pdf"))

Without any text modifications it outputs slightly different text than what I had before, with extra line breaks, but I guess that's from newer versions of various things.

[jbarlow83]

Yes, there are differences between hocr and straight PDF rendering so this workflow is not identical.