ocrs is a Rust library and CLI tool for extracting text from images, also known as OCR (Optical Character Recognition).
The goal is to create a modern OCR engine that:
Under the hood, the library uses neural network models trained in PyTorch, which are then exported to ONNX and executed using the RTen engine. See the models section for more details.
ocrs is currently in an early preview. Expect more errors than commercial OCR engines.
ocrs currently recognizes the Latin alphabet only (eg. English). Support for more languages is planned.
To install the CLI tool, you will first need Rust and Cargo installed. Then run:
$ cargo install ocrs-cli
To extract text from an image, run:
$ ocrs image.png
When the tool is run for the first time, it will download the required models
automatically and store them in ~/.cache/ocrs
.
Extract text from an image and write to content.txt
:
$ ocrs image.png -o content.txt
Extract text and layout information from the image in JSON format:
$ ocrs image.png --json -o content.json
Annotate an image to show the location of detected words and lines:
$ ocrs image.png --png -o annotated.png
See the ocrs crate README for details on how to use ocrs as a Rust library.
ocrs uses neural network models written in PyTorch. See the ocrs-models repository for more details about the models and datasets, as well as tools for training custom models. These models are also available in ONNX format for use with other machine learning runtimes.
To build and run the ocrs library and CLI tool locally you will need a recent stable Rust version installed. Then run:
git clone https://github.com/robertknight/ocrs.git
cd ocrs
cargo run -p ocrs-cli -r -- image.png
Ocrs has unit tests for the code that runs before and after ML model processing, plus E2E tests which exercise the whole pipeline, including models.
After making changes to the code, run unit tests and lint checks with:
make check
You can also run standard commands like cargo test
directly.
Run the E2E tests with:
make test-e2e
For details of how the ML models are evaluated, see the ocrs-models repository.