- Home
- OCR software
- docTR
docTR
docTR · founded 2021
What it is
docTR is a free OCR library for locating and recognizing text in documents, intended for document automation and research. Its pretrained predictors use a two-stage process: they detect text and then recognize it. A separate layout predictor identifies regions such as tables, figures, and headers. The library reads PDFs and images, as well as web pages when the optional HTML extra is installed, and can return plain text or a nested, JSON-serializable dictionary. The project describes its predictors as optimized for CPU and GPU inference. Users can train custom detection, recognition, layout, and table-structure models if pretrained options do not fit their needs. Models can be loaded from and shared through the Hugging Face Hub. Deployment options include a minimal REST API template and a browser demo. Installation is available through pip, Git, or official Docker images, and requires Python 3.11 or higher. Pretrained weights download on first use and are cached locally. The project is licensed under Apache License 2.0.
For and against
For
- Pretrained predictors detect and recognize document text.
- Layout prediction identifies tables, figures, and headers.
- Reads PDFs, images, and optional-extra web pages.
- Supports CPU and GPU inference.
- Offers custom model training and REST API deployment.
Against
- Requires Python 3.11 or higher.
- HTML input requires the optional HTML extra.
- AWS Lambda requires multiprocessing to be disabled and cache moved to /tmp.
The verdict
docTR combines text recognition with document layout analysis and supports custom model training. Its Python requirement and AWS Lambda configuration notes are worth reviewing before deployment.
Who it suits
docTR suits developers and researchers automating document text extraction or building OCR workflows. It may also suit users who need layout detection or want to train custom models.
Plans and prices
docTR open-source Python library
Python 3.11 or higher · install with pip, Git, or Docker
Read 5 Oct 2026 · mindee.github.io
On the record
- Customization
- Users can train custom detection, recognition, layout, and table structure models when pretrained models do not meet their needs.mindee.github.io · read 5 Oct 2026
- Deployment
- The documentation offers templates for a browser demo and API deployment, and documents running docTR on AWS Lambda.mindee.github.io · read 5 Oct 2026
- Extra modules
- The contrib module includes an artefact detector for items such as logos, QR codes, and barcodes, and requires onnxruntime.mindee.github.io · read 5 Oct 2026
- Inference
- The project describes its predictors as optimized for inference on both CPU and GPU.mindee.github.io · read 5 Oct 2026
- Input formats
- docTR can read PDFs, images, and web pages, with HTML input requiring the optional html extra.mindee.github.io · read 5 Oct 2026
- Inputs and outputs
- The quickstart shows loading PDFs, images, and web pages and exporting results as plain text or a JSON-serializable dictionary.mindee.github.io · read 5 Oct 2026
- Supported inputs
- bothmindee.github.io · read 23 Sep 2026
- Installation
- The library can be installed with pip or from Git, and official Docker images are available from GitHub Container Registry.mindee.github.io · read 5 Oct 2026
- Installation requirement
- The library requires Python 3.11 or higher and can be installed from pip, Git, or an official Docker image.mindee.github.io · read 5 Oct 2026
- Integrations
- The documentation describes loading and sharing models through the Hugging Face Hub.mindee.github.io · read 5 Oct 2026
- Layout analysis
- A layout predictor detects document regions such as tables, figures, and headers.mindee.github.io · read 5 Oct 2026
- License
- The documentation's source code identifies the program as licensed under Apache License 2.0.mindee.github.io · read 5 Oct 2026
- Maker
- The project documentation says docTR is actively maintained by Mindee.mindee.github.io · read 5 Oct 2026
- OCR features
- It provides pretrained two-stage text detection and recognition predictors and a layout analysis predictor for regions such as tables, figures, and headers.mindee.github.io · read 5 Oct 2026
- OCR pipeline
- Its pretrained OCR predictors use a two-stage text detection and recognition pipeline.mindee.github.io · read 5 Oct 2026
- Output
- OCR results can be rendered as plain text or exported as a JSON-serialisable nested dictionary.mindee.github.io · read 5 Oct 2026
- Primary platform
- desktopmindee.github.io · read 23 Sep 2026
- Pretrained weights
- Pretrained weights are downloaded on first use and cached locally for later calls.mindee.github.io · read 5 Oct 2026
