docTR: Document Text Recognition

State-of-the-art Optical Character Recognition made seamless & accessible to anyone, powered by TensorFlow 2 & PyTorch

DocTR provides an easy and powerful way to extract valuable information from your documents:

• 🧾 for automation: seamlessly process documents for Natural Language Understanding tasks: we provide OCR predictors to parse textual information (localize and identify each word) from your documents.

• 👩‍🔬 for research: quickly compare your own architectures speed & performances with state-of-art models on public datasets.

Main Features

• 🤖 Robust 2-stage (detection + recognition) OCR predictors with pretrained parameters

• ⚡ User-friendly, 3 lines of code to load a document and extract text with a predictor

• 🚀 State-of-the-art performance on public document datasets, comparable with GoogleVision/AWS Textract

• ⚡ Optimized for inference speed on both CPU & GPU

• 🐦 Light package, minimal dependencies

• 🛠️ Actively maintained by Mindee

• 🏭 Easy integration (available templates for browser demo & API deployment)

Model zoo

Text detection models

• DBNet from “Real-time Scene Text Detection with Differentiable Binarization”

• LinkNet from “LinkNet: Exploiting Encoder Representations for Efficient Semantic Segmentation”

• FAST from “FAST: Faster Arbitrarily-Shaped Text Detector with Minimalist Kernel Representation”

Text recognition models

• SAR from “Show, Attend and Read: A Simple and Strong Baseline for Irregular Text Recognition”

• CRNN from “An End-to-End Trainable Neural Network for Image-based Sequence Recognition and Its Application to Scene Text Recognition”

• MASTER from “MASTER: Multi-Aspect Non-local Network for Scene Text Recognition”

• ViTSTR from “Vision Transformer for Fast and Efficient Scene Text Recognition”

• PARSeq from “Scene Text Recognition with Permuted Autoregressive Sequence Models”

Supported datasets

• FUNSD from “FUNSD: A Dataset for Form Understanding in Noisy Scanned Documents”.

• CORD from “CORD: A Consolidated Receipt Dataset forPost-OCR Parsing”.

• SROIE from ICDAR 2019.

• IIIT-5k from CVIT.

• Street View Text from “End-to-End Scene Text Recognition”.

• SynthText from Visual Geometry Group.

• SVHN from “Reading Digits in Natural Images with Unsupervised Feature Learning”.

• IC03 from ICDAR 2003.

• IC13 from ICDAR 2013.

• IMGUR5K from “TextStyleBrush: Transfer of Text Aesthetics from a Single Example”.

• MJSynth from “Synthetic Data and Artificial Neural Networks for Natural Scene Text Recognition”.

• IIITHWS from “Generating Synthetic Data for Text Recognition”.

• WILDRECEIPT from “Spatial Dual-Modality Graph Reasoning for Key Information Extraction”.