Ico_docTR

docTR

Star

Open-source python document understanding library for developers and data scientists

See on Github
Talk to an expert
This is a preview of docTR

Trainable deep learning OCR enabling the most advanced document understanding use cases

this is a docTR illustration

State-of-the-art

Benefit from the latest computer vision breakthroughs to solve the most complex document processing use cases.

Open source

Build a tailor-made OCR capability that can be hosted in your environment to comply with your data privacy policy.

Trainable

Achieve high extraction performances at scale on US, Europe, or any latin alphabet receipts, from various industries and sectors.

A fully packaged document understanding library for developers and data scientists

Pretrained OCR

Plug and play python OCR trained on millions of latin alphabet documents

See docs

End-to-end-Pipeline

Two-stages OCR pipeline using text detection and recognition algorithms

See more

Training

Text detection and recognition training scripts for PyTorch and TensorFlow

See references

Public Datasets

Built-in support for the most famous OCR challenges public datasets

See datasets

Artefact Detection

Detection algorithms for QR codes, bar codes, signatures, faces...

See docs

Benchmark

Recall, precision and FPS benchmark between different models

See benchmarks

TensorFlow JS

OCR inference in web browser powered by TFJS

See demo app

Local Demo App

Local demo UI generator powered by Streamlit

See docs

Model Compression

Half-precision and quantization support for model optimization

See docs
See full documentation

A fully trainable two-stage approach to OCR

Achieve maximum accuracy by training both the text detection and recognition layers for your specific problem

docTR OCR

Detection models

docTR images

Recognition models:

This is a silly emoji

See live demo on Hugging Face