https://github.com/robertknight/ocrs-models
PyTorch models for the ocrs OCR engine
https://github.com/robertknight/ocrs-models
ocr
Last synced: 2 months ago
JSON representation
PyTorch models for the ocrs OCR engine
- Host: GitHub
- URL: https://github.com/robertknight/ocrs-models
- Owner: robertknight
- Created: 2022-06-19T11:46:07.000Z (almost 3 years ago)
- Default Branch: main
- Last Pushed: 2024-08-20T18:04:03.000Z (9 months ago)
- Last Synced: 2025-02-08T05:11:13.643Z (3 months ago)
- Topics: ocr
- Language: Python
- Homepage:
- Size: 676 KB
- Stars: 62
- Watchers: 2
- Forks: 9
- Open Issues: 9
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# ocrs-models
This project contains tools for training PyTorch models for use with the
[**Ocrs**](https://github.com/robertknight/ocrs/) OCR engine.
## About the models
The ocrs engine splits text detection and recognition into three phases, each
of which corresponds to a different model in this repository:
1. **Text detection**: This is a semantic segmentation model which classifies
each pixel in a greyscale input image as text/non-text. Consumers then
post-process clusters of text pixels to get oriented bounding boxes for
words.
2. **Layout analysis (VERY WIP)**: This is a graph model which takes word
bounding boxes as input nodes and classifies each node's relation to nearby
nodes (eg. start / middle / end of line)
3. **Text recognition**: This is a CRNN model that takes a greyscale image of a
text line as input and returns a sequence of characters.
All models can be exported to ONNX for downstream use.
## Datasets
The models are trained exclusively on datasets which are a) open and b) have non-restrictive licenses. This currently includes:
- [HierText](https://github.com/google-research-datasets/hiertext) (CC-BY-SA 4.0)
## Pre-trained models
Pre-trained models are available from [Hugging
Face](https://huggingface.co/robertknight/ocrs) as PyTorch checkpoints,
[ONNX](https://onnx.ai) and [RTen](https://github.com/robertknight/rten) models.
## Training custom models
See the [Training guide](docs/training.md) for a walk-through of the process to
train models from scratch or fine-tune existing models.