YoloCR is a convenient OCR script.

README_EN.md 6.2KB

YoloCR

For noobs

Install the requirements with the Debian installation script. You can install Debian Stretch inside a virtual machine like Virtualbox.

Requirements

Global Requirements for all the OS.

Global Requirements

Unix/Linux Requirements

  • Tesseract-OCR (we recommend version 3.03+)
    • and install the data corresponding to the languages you want to OCR
    • Imagemagick is required if you use LSTM engine
  • links
  • sxiv (Simple X Image Viewer) (GUI mode only)
  • xdotool (Linux only, GUI mode only)
  • parallel (GNU Parallel)

Note: most of these package, with the exception of all the plugins for vapoursynth, are available as official package for your distro.

For Debian 9, all the requirements can be installed with the YoloDebInstallation script : sh YoloDebInstallation.sh eng-only

For Ubuntu, vapoursynth, vapoursynth-editor and vapoursynth-extra-plugins (to install all the mandatory plugins above) are available through this ppa: ppa:djcj/vapoursynth

Windows Requirements

  • Cygwin. During the install, activate:
    • bc
    • gnupg
    • links
    • make
    • perl
    • wget
    • tesseract-ocr
    • tesseract-ocr-eng

You can use ABBYY FineReader instead of Tesseract.

  • Install GNU Parallel from the Cygwin terminal:
    • wget -O - pi.dk/3 | bash
    • if [ -f ~/bin/parallel ]; then mv ~/bin/parallel /usr/local/bin/; fi

Note: Cygwin terminal usage here → https://help.ubuntu.com/community/UsingTheTerminal

C drive path is "/cygdrive/c".

Scripts have to be used within Cygwin terminal.

If you use Windows 10, you can install BashOnWindows instead of Cygwin.

How to use?

Help for determining the parameters for the YoloCR.vpy file

Determine the Resize parameters.

Resize is very helpful to locate the subtitles.

  1. open YoloResize.vpy in Vapoursynth Editor.
  2. Change this value:
    • FichierSource is the path of the video to OCR.
    • DimensionCropbox allows you to limit the OCR zone.
    • HauteurCropBox allows you to define the height of the Cropbox's bottom border.

Note that theses two parameters have to be adjusted before upscale.

You can then change Supersampling parameter to -1 and verify that your subtitles aren't eated by the white borderline by using F5.

Determine the threshold of the subtitles

It's to improve the OCR-process and the subtitles detection.

  1. Open YoloSeuil.vpy in Vapoursynth editor.
  2. Report FichierSource, DimensionCropBox and HauteurCropBox you have defined in the Resize file.
  3. Choose the fitting ModeS. 'L' if you want to define a white or black threshold, succesively 'R', 'G' and 'B' otherwise.
  4. Adjust the Threshold with the help of the "Color Panel" found in the F5 window.

You must to do this two times if you are using ModeS=L:

  • in the first case, the Inline threshold will be the minimum.
  • in the second case, the Outline threshold will be the maximum.

You can then change Seuil paremeter to the values previously found.

  • in the first case, the subtitles must remain completely visible. The highest the value, the better.
  • in the second case, the Outline must remain completely black. The lowest the value, the better.

Filter the video

  1. Edit the first lines in YoloCR.vpy thanks to the two previos steps (and the previous file YoloSeuil.vpy).

    • SeuilI = the inline threshold value (decrease it if it improves the clarity of the letters)
    • SeuilO = the outline threshold value (increase it if some letters got erased)
  2. Then filter it: vspipe -y YoloCR.vpy - | ffmpeg -i - -c:v mpeg4 -qscale:v 3 -y nameOftheVideoOutput.mp4

Be careful: your must use a different name for your FichierSource in the previous files and nameOftheVideoOutput.mp4 for the output of the ffmpeg command.

You now have an OCR-isable video and scenechange informations.

OCR the video

Then you can OCR the video: ./YoloCR.sh nameOftheVideoOutput.mp4

The nameOftheVideoOutput.mp4 must be the same than the output of the ffmpeg command.

You can use YoloTime.sh instead of YoloCR.sh if you only want the Timing of the subtitles.

Now it's Checking time :D

Serial use of YoloCR

  1. Make sure that sxiv isn't installed.
  2. Make sure that YoloCR directory includes the video files you want to OCR and only theses.
  3. Comment the first line of YoloCR.vpy. ("FichierSource" becomes "#FichierSource".)
  4. Move to the YoloCR directory and use this bash command:
    • for file in *.mp4; do filef="${file%.*}_filtered.mp4"; vspipe -y --arg FichierSource="$file" YoloCR.vpy - | ffmpeg -i - -c:v mpeg4 -qscale:v 3 -y "$filef"; ./YoloCR.sh "$filef"; done

"*.mp4" means that all files having the mp4 extension will be processed. Read about bash regex if you want more flexibility.

Known bugs

  • Tesseract's LSTM engine produce a lower quality OCR (such as a worse italics detection).
  • Cygwin (Windows), when you run YoloCR.sh for the first time.
    • Signal SIGCHLD received, but no signal handler set.
    • YoloCR will run without errors the next times.
  • Babun (Windows), you will have errors when trying to run YoloCR.sh.
    • Use Cygwin instead.

Contact: https://discord.gg/USD6wK3