ocrs

ocrs is an OCR engine written in Rust. It extracts text, including layout information, from images. It uses the RTen machine learning runtime.

ocrs uses neural network models written in PyTorch. See the ocrs-models repository for more details and tools for training custom models. These models are also available in ONNX format for use with other machine learning runtimes.

Background and status

The original goal of this project was to create a more modern alternative to Tesseract, which is still (!) the de-facto open-source OCR engine. See https://github.com/robertknight/tesseract-wasm/issues/87 for context. Some ways in which this project aims to improve upon Tesseract are:

• Reducing the need to manually clean up and pre-process images before feeding them into engine. This is achieved by replacing manually written components for text detection and, in future, layout analysis, with machine learning models.
• Using modern tools (eg. PyTorch) for training. Together with use of open datasets this should make it easier to recreate and customize models
• Making the models available in portable formats (ONNX) which can be used with many different runtimes
• Better support for Unicode and large character sets.

ocrs is currently in an early preview. Expect more errors than commercial OCR engines. The layout analysis phase is still uses manually written code which can be brittle.

Command-line usage

This library is wrapped in a CLI tool in the ocrs-cli crate, which you can use as follows:

$ cargo install ocrs-cli
$ ocrs image.png

See the ocrs-cli README for more details.

Library usage

See examples/hello_ocr.rs for a minimal example of using this library in a Rust application.

cd examples/

# Download models in .rten format.
./download-models.sh

# Run OCR on an image and print the extracted text.
cargo run -r --example hello_ocr rust-book.jpg