OCRbot is a Mastodon (and compatible, such as Pleroma) bot that uses OCR to automatically generate text descriptions of images. It reads the image and outputs what it thinks is the text contained, using tesseract. It requires Python 3.
A working version of OCRbot can be found here. I'll always run the latest version, and the instance hosting it has a much higher character limit than Mastodon's default of 500 or Pleroma's default of 5000 (specifically, it has a limit of 65535 chars). If you'd like to run your own version that works differently (for example, maybe you want a version that works with Japanese text instead, or a version that uses a different OCR engine), then you're free to fork and modify!
sudo apt install tesseract-ocr
- Arch Linux:
sudo pacman -Syu tesseract tesseract-data-eng
What about the higher quality data packs?
According to tesseract's GitHub wiki:
Most users will want
tessdata_fastand that is what will be shipped as part of Linux distributions.
tessdata_bestis for people willing to trade a lot of speed for slightly better accuracy.
With the setup I'm using, I really can't afford to use much more processing power just for "slightly better accuracy". Additionally, most Linux distributions only ship the
fast set, which makes obtaining the
best set an extra step. I haven't tested how much better
fast is than
best, but given that distributions like Arch and Debian deemed
fast good enough to be the only package available, I'm inclinded to suspect that this is a case of diminishing returns, and that it's not worth the extra effort of obtaining these files, as well as the extra CPU usage required to use them.
git clone https://github.com/Lynnesbian/OCRbot/ cd OCRbot pip3 install -r requirements.txt
login.py, and answer the interactive prompts. To further configure OCRbot, open
config.json and edit the settings as you wish. Here is an explanation of the options in the config file:
|site||The instance your bot will log in to and post from.||https://botsin.space|
|cw||The content warning (aka subject) OCRbot will apply to non-error posts.||OCR output|
|ocr_threads||The amount of CPU threads to use for running tesseract.||8|
|char_limit||The maximum length of an OCRbot post. This limit does not apply to errors. I recommend setting it to your instance's character limit.||500|
|default_language||The language you'd like OCRbot to default to.||eng|
|admin||The account you'd like OCRbot to tell users to report errors to. If you format it as
Most importantly, make sure you choose the instance your OCRbot account will post from. Then, run
login.py and log in with the account you'd like OCRbot to post from. Finally, run
service.py and OCRbot will take care of the rest.
Running OCRbot as a Service
You can use something like systemd or SysVinit to manage running
service.py for you. I've provided an example systemd script here. It requires some editing (to specify where your OCRbot folder is located, and to specify the user to run it as -- don't run it as root), and then you simply need to
sudo mv systemd-example.service /etc/systemd/system/OCRbot.service sudo systemctl start OCRbot
You may also
sudo systemctl enable OCRbot if you want it to start on boot.
OCRbot is not designed to replace captions, merely to supplement them. It works best with plain black text on a plain white background. For example, PDF scans and Wikipedia screenshots will work great, but handwriting and distorted text won't. It will almost always fail (at least partially) for:
- Anything using handwriting, especially cursive
- Text that is skewed, rotated, stretched, or otherwise distorted
- Heavily compressed images
- Text with layouts more complex than this page
- Unusual fonts
- Videos and GIFs
- Very large or very small text
- Text on a complex background
- Vertical text in a language where text is usually written horizontally, and vice versa
Additionally, OCRbot may sometimes fail to find images due to federation issues. This is a known issue, and OCRbot does its best to work around it, but due to the way federation works, it will never be completely solved.
If you can caption images yourself, I recommend doing so. If, for example, you are posting a screenshot of a tweet, you'll get better results from copy and pasting the text of the tweet as a caption rather than relying on OCRbot. OCRbot's image recognition is far from perfect and never will be.
Support for Other Languages
OCRbot is capable of supporting languages other than English. If you are using Debian, you can install the package
tesseract-ocr-all to install all language files, or install individual ones, such as
tesseract-ocr-jpn for Japanese. Arch also provides these packages, with slightly different names (e.g.
tesseract-data-jpn), although it does not have an "all" package.
While invoking OCRbot, you may encounter a number of errors. While the text is somewhat self explanatory, this section exists to provide further elaboration on their meaning.
|Couldn't find any media.||OCRbot was unable to find any posts that had images attached, usually because your post or the one you're replying to doesn't have any images. It could also happen if the images have failed to federate properly.|
|Failed to find post containing image. This may be a federation issue, or you may have tagged OCRbot in a conversation without an image. Please note that OCRbot can only see images in the post you are directly replying to, and can't see images that are provided as URLs rather than attachments.||(see above)|
|Specified post has no images.||OCRbot successfully loaded the post, and was unable to find any images. This means that your post or the one you're replying to definitely doesn't have any images.|
|Couldn't read the specified image(s), sorry! OCRbot works best with plain black text on a plain white background. Here is some information about what it can and can't do: https://github.com/Lynnesbian/OCRbot/blob/master/README.md#caveats||OCRbot couldn't read your image. In the same way that you would have trouble reading a blurry photo of text, OCRbot has trouble with things such as cursive handwriting, rotated text, strange fonts, and more. Check the caveats section on this page for more information.|
|Tesseract returned no text.||Either your image contains no text, or tesseract (the program OCRbot relies on) encountered an error and returned nothing.|
|Failed to read image. Download may have timed out.||If any one image takes longer than 30 seconds to download, OCRbot gives up on it. This can occur because a server is down, experiencing slowdown, or most likely, because wasabi is down.|
|Failed to run tesseract. Specified language was: [language]||Tesseract, the program OCRbot relies on to provide captions, crashed.|
Copyright (C)2019 Lynnesbian (https://fedi.lynnesbian.space/@lynnesbian)
This program is free software: you can redistribute it and/or modify it under the terms of the GNU Affero General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version.
This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU Affero General Public License for more details.
You should have received a copy of the GNU Affero General Public License along with this program. If not, see https://www.gnu.org/licenses/.
The full text of the license is provided under the name "LICENSE" in the root directory of this repository.