How to extract text with OCR from a PDF on Linux?

Question

How do I extract text from a PDF that wasn't built with an index? It's all text, but I can't search or select anything. I'm running Kubuntu, and Okular doesn't have this feature.

Solution 1

I have had success with the BSD-licensed Linux port of Cuneiform OCR system.

No binary packages seem to be available, so you need to build it from source. Be sure to have the ImageMagick C++ libraries installed to have support for essentially any input image format (otherwise it will only accept BMP).

While it appears to be essentially undocumented apart from a brief README file, I've found the OCR results quite good. The nice thing about it is that it can output position information for the OCR text in hOCR format, so that it becomes possible to put the text back in in the correct position in a hidden layer of a PDF file. This way you can create "searchable" PDFs from which you can copy text.

I have used hocr2pdf to recreate PDFs out of the original image-only PDFs and OCR results. Sadly, the program does not appear to support creating multi-page PDFs, so you might have to create a script to handle them:

#!/bin/bash
# Run OCR on a multi-page PDF file and create a new pdf with the
# extracted text in hidden layer. Requires cuneiform, hocr2pdf, gs.
# Usage: ./dwim.sh input.pdf output.pdf

set -e

input="$1"
output="$2"

tmpdir="$(mktemp -d)"

# extract images of the pages (note: resolution hard-coded)
gs -SDEVICE=tiffg4 -r300x300 -sOutputFile="$tmpdir/page-%04d.tiff" -dNOPAUSE -dBATCH -- "$input"

# OCR each page individually and convert into PDF
for page in "$tmpdir"/page-*.tiff
do
    base="${page%.tiff}"
    cuneiform -f hocr -o "$base.html" "$page"
    hocr2pdf -i "$page" -o "$base.pdf" < "$base.html"
done

# combine the pages into one PDF
gs -q -dNOPAUSE -dBATCH -sDEVICE=pdfwrite -sOutputFile="$output" "$tmpdir"/page-*.pdf

rm -rf -- "$tmpdir"

Please note that the above script is very rudimentary. For example, it does not retain any PDF metadata.

Solution 2

Google docs will now use OCR to convert your uploaded image/pdf documents to text. I have had good success with it.

They are using the OCR system that is used for the gigantic Google Books project.

However, it must be noted that only PDFs to a size of 2 MB will be accepted for processing.

Update
1. To try it out, upload a <2MB pdf to google docs from a web browser.
2. Right click on the uploaded document and click "Open with Google Docs".
...Google Docs will convert to text and output to a new file with same name but Google Docs type in same folder.

Solution 3

Best and easyest way out there is to use pypdfocr it doesn't change the pdf

pypdfocr your_document.pdf

At the end you will have another your_document_ocr.pdf the way you want it with searchable text. The app doesn't change the quality of the image. Increases the size of the file a bit by adding the overlay text.

Update 3rd november 2018:

pypdfocr is no longer supported since 2016 and I noticed some problems due to not being mentained. ocrmypdf(module) does a symiliar job and can be used like this:

ocrmypdf in.pdf out.pdf

To install:

pip install ocrmypdf

or

apt install ocrmypdf

Solution 4

Geza Kovacs has made an Ubuntu package that is basically a script using hocr2pdf as Jukka suggested, but makes things a bit faster to setup.

From Geza's Ubuntu forum post with details on the package...

Adding the repository and installing in Ubuntu

sudo add-apt-repository ppa:gezakovacs/pdfocr
sudo apt-get update
sudo apt-get install pdfocr

Running ocr on a file

pdfocr -i input.pdf -o output.pdf

GitHub repository for the code https://github.com/gkovacs/pdfocr/

Solution 5

PDFBeads works well for me. This thread “Convert Scanned Images to a Single PDF File” got me up and running. For a b&w book scan, you need to:

1. Create an image for every page of the PDF; either of the gs examples above should work
2. Generate hOCR output for each page; I used tesseract (but note that Cuneiform seems to work better).
3. Move the images and the hOCR files to a new folder; the filenames must correspond, so file001.tif needs file001.html, file002.tif file002.html, etc.

4. In the new folder, run

   pdfbeads * > ../Output.pdf
   

This will put the collated, OCR'd PDF in the parent directory.

View more solutions

Comments

[Gökhan Sever]

Any idea to improve this script to add spell-checking stage to correct errors in recognition step?

[Jukka Matilainen]

@Gökhan Sever, do you mean adding interactive spell-checking where the user is prompted for replacement for misspelled/unknown words? I think you could do that by adding something like aspell check --mode=html "$base.html" in the script right after running cuneiform.

[Pitto]

The answer is not really Ubuntu-specific but I want to really thank you: BRILLIANT solution! :)

[Keks Dose]

Small correction: The line for tesseract at least for other languages than English, here e.g. German ( = deu ) is: ` tesseract "$page" "$base" -l deu hocr ` (of course you have to remove the ` `).

[Admin]

As I had problems with not so accurate pdfs I changed the engine in gs from "tiff4" to "tiffgray" - and the result was very good: gs -SDEVICE=tiffgray -r300x300 -sOutputFile="$tmpdir/page-%04d.tiff" -dNOPAUSE -dBATCH -- "$input"

[fixer1234]

I found ABBYY OCR to be pretty pitiful, one of the least capable programs I've tried. It might be adequate with a really clean image of standard font text of typical body text size, with no mixed fonts, mixed sizes, complex layout, graphics, lines, etc.

[fixer1234]

If what you need isn't covered in other answers here, the best thing to do is ask your own question. That will get it exposure to a lot of eyes.

[Wikunia]

@GökhanSever I'll get this error: Tesseract Open Source OCR Engine v3.03 with Leptonica OSD: Weak margin (0.00) for 571 blob text block, but using orientation anyway: 0 /usr/bin/pdf2text: line 23: /tmp/tmp.XksXutALLp/page-0001.html: No such file or directory when I use your version. Any idea what I'm doing wrong?

[fixer1234]

This post states that the product can do it, which is a helpful hint that should be posted as a comment. It doesn't explain how to actually solve the problem, which is what answers should do. Can you expand your answer so that someone can see how to do the solution?

[Scanner.js Receipt Invoice OCR]

Thanks @fixer1234, I've edited it to include the command.

[Gaurav]

This was really helpful :) I uploaded a 50 MB file yesterday and it worked. Looks like they've increased the size limit.

[David Milovich]

@Wikunia change $base.html to $base.hocr