mindee/doctr is docTR, a deep-learning-based OCR library written in Python. The latest release published on GitHub is v1.1.0, and the license is Apache-2.0. Among OCR-related tasks it covers document text recognition — that is, locating the characters in a document or image and reading them.
Extracting text from a document proceeds in two stages. A detector finds where the words are in the document, and a recognizer reads the characters at those positions. The output carries the characters together with their structural positions, and a document layout analysis feature can even recover the overall arrangement. Accepted inputs are PDF, images, and web pages, and the implementation is built on PyTorch. Mindee created the project, and t2k GmbH now handles development and maintenance. The repository was started on January 8, 2021, and has 6,391 stars.
However, what changed in v1.1.0 cannot be confirmed from the publicly available material alone. There are no release notes or changelog, and no benchmarks comparing accuracy or processing speed are presented. Whether a changelog and performance data will be published together is the next point to check. If you are looking for a place to slot OCR into a document processing flow, this is worth examining, but it does not appear to be the stage at which to decide on adoption without performance figures.