ucr 0.2.15

Universal Character Recognizer (UCR): Simple, Intuitive, Extensible, Multi-Lingual OCR engine

            

Universal Character Recognizer (UCR) is an Open Source, Easy to use Python library to build Production Ready OCR applications with its highly Intuitive, Modular & Extensible API design and off-the-shelf Pretrained Models for over 25 languages.

Read UCR Documentation on ucr.docyard.ai

Features • Setup • Usage • Acknowledgement

Demo

For details, click here!

Features

• Supports SOTA Text Detection and Recognition models
• Built on top of Pytorch and Pytorch Lightning
• Supports over 25 languages
• Model Zoo contains 27 Pretrained Models across 25 languages
• Modular Design Language allows Pick and Choose of different components
• Easily extensible with Custom Components and attributes
• Hydra config enables Rapid Prototyping with multiple configurations
• Support for Packaging, Logging and Deployment tools straight out of the box

Note: Some features are still in active development and might not be available.

Setup

Installation

Require python version >= 3.6.2, install with pip (recommended)

1. Prerequisites: Install compatible version of Pytorch and torchvision from official repository.
2. Installation: Install the latest stable version of UCR:

pip install -U ucr

[Optional] Test Installation

Run dummy tests!

ucr test
# Optional: Add -l/--lang='language_id' to test on particular language!
ucr test -l='en_number'

Usage

Workflow

Execution flow of UCR is displayed above. Broadly it can be divided into 4 sub-parts:

1. Input (image/folder path or web address) is fed into the Detection model which outputs bounding box coordinates of all the text boxes.
2. The detected boxes are then checked for Orientation and corrected accordingly.
3. Next, Recognition model runs on the corrected text boxes. It returns bounding box information and OCR output.
4. Lastly, an optional Post Processing module is executed to improve/modify the results.

Quick Start

The following code snippet shows how to get started with UCR library.

from ucr import UCR

# initialization
ocr = UCR(lang="en_number", device="cpu")

# run prediction
result = ocr.predict('input_path', output='output_path')

# for saving annotated image
result = ocr.predict('input_path', output='output_path', save_image=True)

For complete list of arguments, refer Argument List

Model Zoo

A collection of pretrained models for detection, classification and recognition processes is present here !
These models can be useful for out-of-the-box inference on over 25 languages.

Acknowledgement

Substantial part of the UCR library is either inspired or inherited from the PaddleOCR library. Wherever possible the repository has been ported from PaddlePaddle to PyTorch framework including the direct translation of model parameters. Also, a big thanks to Clova AI, for open sourcing their testing script and pretrained models (CRAFT).

License

Apache License 2.0