[OCR integration] Add OnnxTR as possible OCR engine

[felixT2K]

Requested Feature

I noticed that your project already supports multiple OCR integrations (EasyOCR, Tesseract, and RapidOCR/PaddleOCR). I believe OnnxTR would be a valuable addition.

OnnxTR is a core-refactored version of docTR that runs on ONNX Runtime, similar to how RapidOCR is built on PaddleOCR.

What makes OnnxTR/docTR special?

• Modular design: Detection, recognition, and orientation predictors can be used independently.
• Robust to rotation: Handles rotated documents effectively and includes auto-correction.
• Highly customizable: Each predictor has its own configurable EngineConfig.
• Flexible installation options: Supports CPU, GPU, and OpenVINO, with CI testing on Windows, Linux, and macOS (Python 3.10–3.12).
• Model hub integration: Models can be loaded directly from the Hugging Face Hub, e.g.:
  OnnxTR Models

Known Limitation

• The default pre-trained models are trained on French vocabulary. However, one model on the Hugging Face Hub supports Latin Extended characters (covering English, French, German, Spanish, Portuguese, Italian, etc.).

• Ongoing language expansion: Work is in progress to support additional languages such as Russian, Hebrew, Hindi, and more.

References

• docTR Repository
• OnnxTR Repository

Testing

• OnnxTR Demo (Hugging Face Space)
• docTR Demo (Hugging Face Space)

[felixT2K]

@PeterStaar-IBM @dolfim-ibm wdyt ? :)