Powerful OCR system under GNU/Linux for PDF documents managed from command line and with refinement by Vim.

Introduction.

The idea came from reading this article about optical character recognition (OCR) in the GNU/Linux environment from images and PDF, managed from the command line.

Obviously, PDF documents are those scanned from paper original, i.e., not obtained by direct saving of document in digital format. For the latter, no OCR is needed.

The article is very well written and the end result is very good.

I wondered if it would be possible to aggregate all the steps into a single text command.

In this article I report my solution.

Next, then, I added some con RegEx steps by Vim to reformat the raw result of optical recognition.

Again, I tried to combine several separate formulas into a single regular expression.

So the goal is to use one text command and one regular expression with Vim to achieve good optical recognition in GNU/Linux environment.

The installation of components.

The system engine is Tesseract, the installation of which, well described in the reference article, is done as follows:

  • On Debian and derivatives, including Ubuntu and Mint: sudo apt-get install tesseract-ocr.
  • On Fedora: sudo dnf install tesseract.
  • On Arch and derivatives, including Manjaro and EndeavourOS: sudo pacman -S tesseract or yay -S tesseract.

Please note: for the “family” Debian the component is named tesseract-ocr while for other distributions, simply tesseract.

For the installation of additional languages it is necessary to use the system already indicated for each distribution by pointing to one of the language codes indicated on this page, taking into account some semantic differences between the various distributions, as indicated below in the case of the data related to the Italian language:

  • On Debian and derivatives, including Ubuntu and Mint: sudo apt-get install tesseract-ocr-ita.
  • On Fedora: sudo dnf install tesseract-langpack-ita.
  • On Arch and derivatives, including Manjaro and EndeavourOS: sudo pacman -S tesseract-data-ita or yay -S tesseract-data-ita. Arch also makes it possible to install data along with the basic program.

It is also necessary to verify that the pdftoppm software, which is part of the poppler-utils package, is present in the system.

The test is very simple: just type pdftoppm -help. If present, it will respond with the various options available.

If not, you will need to install the whole package:

  • On Debian and derivatives, sudo apt install poppler-utils.
  • On Fedora: sudo dnf popple-utils.
  • On Arch and derivatives: sudo pacman -S poppler or yay -S poppler.

OCR of PDF documents with “tesseract”: description of steps.

This is the outline of the steps to achieve the result:

  • Objective: submit documents in PDF format to optical recognition (OCR).
  • Problem: tesseract operates only on images and not, instead, on PDF documents.
  • Solution: some operations need to be performed on PDF documents before and after optical recognition, as follows:
    • transform PDFs into images, separating any multiple pages into single images,
    • apply OCR software to each individual image resulting in many separate text documents,
    • reassemble the individual text documents to reconstruct the original sequence of the source document.

It sounds tricky, but it really isn’t at all if you understand the specific function of each step.

The single steps.

Here are the steps, assuming the document to be processed is named document.pdf:

  1. pdftoppm -png document.pdf document to transform PDFs into PNG-format images (note the -png option) by splitting each page into a single file with progressive numbering.
  2. for i in document-*png; do tesseract "$i" "text-$i" -l ita; done; creates a loop that applies tesseract to each individual image and generates a text document named text-document- followed by progressive double-digit numbering.
  3. cat text-document* > document.txt, to recompose all the individual text documents generated by optical recognition into a single document.

In step #2 I use the command for i in document-* ... to allow the system to run regardless of the number of documents to be processed.

Everything in one command!

And here is the part that accomplishes the first goal of this article: concentrating all steps in a one command!

I create a file that I name ocr.sh (but, of course, that could have any other name).

I insert the following script into the file:

pdftoppm -png document.pdf document 
for i in document-*png; do tesseract "$i" "text-$i" -l ita; done; 
cat text-document* > document.txt
rm document-*
rm text-document-*

The last two lines, namely:

rm document-*
rm text-document-*

automatically delete scattered files generated during the optical recognition process and no longer needed after merging the text into document.txt.

I make the file executable with chmod u+x ocr.sh, for the owner or chmod a+x ocr.sh for all users.

I launch the command having a PDF document named document.pdf present in the same folder as the executable.

The whole process happens in total automatism: at the end I find the result of the optical recognition in the file document.txt.

In other words: the whole procedure will happen in automatic sequence, from the first to the last formula, and, at the end, we will find a document named document.txt with the optical recognition extracted from the above sequence.

If any error occurs during the process, the script would be interrupted with a message identifying the problem (for example: the lack of language form), so that the problem can be corrected.

And now: Vim with RegEx.

And now we can use the power of Vim with RegEx to make the raw result of optical recognition more readable and manageable.

Specifically:

  1. Eliminate all carriage returns on each line, keeping only the blank lines intact.
  2. Reduce all redundant blank lines between paragraphs to one.
  3. Eliminate redundant white characters within the text.

The fomulas in RegEx for each of the above tasks are as follows:

  1. :%s/.\+\zs\n\ze./ /
  2. :g/^\s*$\n^\s*$/d
  3. :%s/\s\+/ /g

The first search string always finds a positive match in the generated document since a forced carriage return is always added to the end of each line.

The other two search strings may not find any positive match in the generated document and in this case would give the result “pattern not found”.

The above, can be launched in a single command (the second goal!) including all the previous ones, (using the pattern described on this page):

:%s/.\+\zs\n\ze./ / | :g/^\s*$\n^\s*$/d | :%s/\s\+/ /g

Warning: in this case the script would be interrupted if the second or third (as anticipated above) search string did not find a match, that is, respectively, if there were no redundant lines or spaces in the document.

At this point, if the interruption occurred at the second string, one could consider, if necessary, launching the third script separately to clean up the internal spaces in the document.

In Conclusion.

At this point we have achieved our stated goals of obtaining a high-quality optical recognition system in GNU/Linux, managed at the command line with a single command for the optical recognition phase and a single RegEx function in Vim for formatting the result.

Thank you for your attention.