perspectra 0.2.1

Automatically extract and perspectively correct documents in images

Perspectra

Software and corresponding workflow to scan documents and books with as little hardware as possible.

Check out github:adius/awesome-scanning for and extensive list of alternative solutions.

Installation

We recommend to use uv instead of pip to install the package.

uv tool install perspectra

To install from source:

git clone https://github.com/ad-si/Perspectra
cd Perspectra
make install

Usage

Command Line Interface

usage: perspectra [-h] [--debug] {binarize,correct,corners,renumber-pages} ...

options:
  -h, --help            show this help message and exit
  --debug               Render debugging view

subcommands:
  subcommands to handle files and correct photos

  {binarize,correct,corners,renumber-pages}
                        additional help
    binarize            Binarize image
    correct             Pespectively correct and crop photos of documents.
    corners             Returns the corners of the document in the image
                        as [top-left, top-right, bottom-right, bottom-left]
    renumber-pages      Renames the images in a directory according to their page numbers.
                        The assumend layout is `cover -> odd pages -> even pages reversed`

Best Practices for Taking the Photos

Your photos should ideally have following properties:

• Photos with 10 - 20 Mpx
• Contain 1 document
  • Rectangular
  • Pronounced corners
  • Only black content on white or light-colored paper
  • On dark background
  • Maximum of 30° rotation

Camera Settings

# Rule of thumb is the inverse of your focal length,
# but motion blur is pretty much the worst for readable documents,
# therefore use at least half of it and never less than 1/50.
shutter: 1/50 - 1/200 s

# The whole document must be sharp even if you photograph it from an angle.
# Therefore at least 8 f.
aperture: 8-12 f

# Noise is less bad than motion blur => relative high ISO
# Should be the last thing you set:
# As high as necessary as low as possible
iso: 800-6400

When using Tv (Time Value) or Av (Aperture Value) mode use exposure compensation to set lightness value below 0. You really don't want to overexpose your photos as the bright pages are the first thing that clips.

On the other hand, it doesn't matter if you loose background parts because they are to dark.

Generating the Photos from a Video

A good tool for this purpose is PySceneDetect. It's a Python/OpenCV-based scene detection program, using threshold/content analysis on a given video.

For easy installation you can use the docker image

Find good values for threshold:

docker run \
  --rm \
  --volume (pwd):/video \
  handflucht/pyscenedetect
  --input /video/page-turning.mp4 \
  --downscale-factor 2 \
  --detector content \
  --statsfile page-turning-stats.csv

To launch the image run:

docker run \
  --interactive \
  --tty \
  --volume=(pwd):/video \
  --entrypoint=bash \
  handflucht/pyscenedetect

Then run in the shell:

cd /video
scenedetect \
  --input page-turning.mp4 \
  --downscale-factor 2 \
  --detector content \
  --threshold 3 \
  --min-scene-length 80 \
  --save-images

TODO: The correct way to do this: (after https://github.com/Breakthrough/PySceneDetect/issues/45 is implemented)

docker run \
  --rm \
  --volume (pwd):/video \
  handflucht/pyscenedetect \
  --input /video/page-turning.mp4 \
  --downscale-factor 2 \
  --detector content \
  --threshold 3 \
  --min-scene-length 80 \
  --save-images <TODO: path>

Aim for a low threshold and a long minimun scene length. I.e. turn the page really fast and show it for a long time.

TODO

• Calculate aspect ratio of scanned document and apply during perspective transformation
• Make sure besin for watershed algorithm is in no local confinement
• Add white border to sobel image or crop it by 1px in order to correctly handle partly cropped documents
• Check that there were at least 4 corners detected
• Dewarp pages
• Maximize contrast of image before binarizing
• Make algorithms independent from photo sizes
• Limit the kind of objects which get deleted when touching the border (e.g. only elongated objects)
• Better algorithm for documents with rounded corners (e.g. credit cards)
• Mention that file format is infered from file extension
• Spread range after converting to grayscale
• Add a border around the image so that watershed starts to flood the image from all directions and to improve recognition for clipped documents
• Implement https://github.com/scikit-image/scikit-image/issues/2212
• Checkout http://ilastik.org
• Trim images before saving
• Try out https://github.com/andrewssobral/bgslibrary
• Try out https://github.com/Image-Py/imagepy
• Try out https://github.com/WPIRoboticsProjects/GRIP/releases
• Try out https://mybinder.org
• Read https://gilberttanner.com/blog/detectron2-train-a-instance-segmentation-model
• Read https://www.pyimagesearch.com/2018/07/30/opencv-object-tracking/