Automate the workflow around ADF scanning, OCR and PDF creation
adf2pdf - a tool that turns a batch of paper pages into a PDF with a text layer. By default, it detects empty pages (as they may easily occur during duplex scanning) and excludes them from the OCR and the resulting PDF.
For that, it uses Sane's scanimage for the scanning, Tesseract for the optical character recognition (OCR), and the Python packages img2pdf, Pillow (PIL) and PyPDF2 for some image-processing tasks and PDF mangling.
$ adf2pdf contract-xyz.pdf
2017, Georg Sauthoff firstname.lastname@example.org
- Automatic document feed (ADF) support
- Fast empty page detection
- Overlaying of scanning, image processing, OCR and PDF creation to minimize the total runtime
- Fast creation of small PDFs using the fine img2pdf package
- Only use of safe compression methods, i.e. no error-prone symbol segmentation style compression like JBIG2 or JB2 that is used in Xerox photocopiers and the DjVu format.
Adf2pdf can be directly installed with
$ pip3 install --user adf2pdf
$ pip3 install adf2pdf
See also the PyPI adf2pdf project page.
Alternatively, the Python file
adf2pdf.py can be directly
executed in a cloned repository, e.g.:
$ ./adf2pdf.py report.pdf
In addition to that, one can install the development version from a cloned work-tree like this:
$ pip3 install --user .
A scanner with automatic document feed (ADF) that is supported by Sane. For example, the Fujitsu ScanSnap S1500 works well. That model supports duplex scanning, which is quite convenient.
Running adf2pdf for a 7 page example document takes 150 seconds
on an i7-6600U (Intel Skylake, 4 cores) CPU (using the ADF of the
Fujitsu ScanSnap S1500). With the defaults, adf2pdf calls
scanimage for duplex scanning into 600 dpi lineart (black and
white) images. In this example, 6 pages are empty and thus
automatically excluded, i.e. the resulting PDF then just contains
The resulting PDF contains a text layer from the OCR such that one can search and copy'n'paste some text. It is 1.1 MiB big, i.e. a page is stored in 132 KiB, on average.
The script assumes Tesseract version 4, by default. Version 3 can be used as well, but the new neural network system in Tesseract 4 just performs magnitudes better than the old OCR model. Tesseract 4.0.0 was released in late 2018, thus, distributions released in that time frame may still just include version 3 in their repositories (e.g. Fedora 29 while Fedora 30 features version 4). Since version 4 is so much better at OCR I can't recommend it enough over the stable version 3.
Tesseract 4 notes (in case you need to build it from the sources):
- Build instructions - warning: if you miss the
autoconf-archivedependency you'll get weird autoconf error messages
- Data files - you need the training data for your languages of choice and the OSD data
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.