I've had a few repos to look through, but those just did not do the trick in sorting out a bunch of scanned paper books with the chess diagrams. First, I had to write the script to separate the diagrams from the rest of the page, then another one to "condition" those extracted diagrams so the recognition actually provides the acceptable results.
- https://github.com/metterklume/chessputzer
- https://github.com/linrock/chessboard-recognizer
- https://github.com/Elucidation/tensorflow_chessbot
I'd say I used these three as an inspiration, but ended up writing my completely own code with similar purpose and probably based on the same ideas. Hopefully my implementation actually does the trick and given a bunch of the book page scans in relatively poor condition, returns accurate representation of the chess position ignoring text and other formatting printed on the page.