docTR (Document Text Recognition) is an open-source deep-learning OCR library originally created by Mindee. It has about 6,400 stars on GitHub and remained actively developed in 2026, with maintenance now led by t2k GmbH.
docTR follows the classic two-stage OCR architecture and does both stages well: a detection model first locates text regions โ with DBNet, LinkNet, and FAST among the choices โ and a recognition model then reads them, with options such as CRNN, SAR, and PARSeq. Implementations are provided for both PyTorch and TensorFlow 2, and the library goes beyond raw text extraction with predictors for layout analysis, table structure recognition, and key information extraction. For deployment, docTR ships Docker images and a minimal API, so a trained pipeline can be served in production without much glue code. The result is an OCR toolkit that is equally at home in a research notebook and a document-processing backend. The project sits in a broader open-source document-understanding ecosystem and is frequently used as the OCR stage in larger pipelines for invoice processing, form extraction, and archive digitization. Its modular design also makes it a practical testbed for comparing detection and recognition architectures on exactly the same documents.
docTR is free and open source under the Apache-2.0 license. Mindee’s commercial document-processing API is a separate product; this library itself costs nothing.
Developers and data teams adding OCR to invoices, forms, archives, and other document pipelines who want modern deep-learning models without a paid API dependency.