docling-ocr-onnxtr 0.1.2

Onnx Text Recognition (OnnxTR) OCR plugin for docling

The docling-OCR-OnnxTR repository provides a plugin that integrates the OnnxTR OCR engine into the Docling framework, enhancing document processing capabilities with efficient and accurate text recognition.

Key Features:

• Seamless Integration: Easily incorporate OnnxTR's OCR functionalities into your Docling workflows for improved document parsing and analysis.

• Optimized Performance: Leverages OnnxTR's lightweight architecture to deliver faster inference times and reduced resource consumption compared to traditional OCR engines.

• Flexible Deployment: Supports various hardware configurations, including CPU, GPU, and OpenVINO, allowing you to choose the best setup for your needs.

Installation:

To install the plugin, use one of the following commands based on your hardware:

For GPU support please take a look at: ONNX Runtime.

• Prerequisites: CUDA & cuDNN needs to be installed before Version table.

# For CPU
pip install docling-ocr-onnxtr[cpu]
# For Nvidia GPU
pip install docling-ocr-onnxtr[gpu]
# For Intel GPU / Integrated Graphics
pip install docling-ocr-onnxtr[openvino]

# Headless mode (no GUI)
# For CPU
pip install docling-ocr-onnxtr[cpu-headless]
# For Nvidia GPU
pip install docling-ocr-onnxtr[gpu-headless]
# For Intel GPU / Integrated Graphics
pip install docling-ocr-onnxtr[openvino-headless]

By integrating OnnxTR with Docling, users can achieve more efficient and accurate OCR results, enhancing the overall document processing experience.

Usage

from docling.datamodel.pipeline_options import PdfPipelineOptions
from docling.document_converter import (
    ConversionResult,
    DocumentConverter,
    InputFormat,
    PdfFormatOption,
)
from docling_ocr_onnxtr import OnnxtrOcrOptions


def main():
    # Source document to convert
    source = "https://arxiv.org/pdf/2408.09869v4"

    # Available detection & recognition models can be found at
    # https://github.com/felixdittrich92/OnnxTR

    # Or you choose a model from Hugging Face Hub
    # Collection: https://huggingface.co/collections/Felix92/onnxtr-66bf213a9f88f7346c90e842

    ocr_options = OnnxtrOcrOptions(
        # Text detection model
        det_arch="db_mobilenet_v3_large",
        # Text recognition model - from Hugging Face Hub
        reco_arch="Felix92/onnxtr-parseq-multilingual-v1",
        # This can be set to `True` to auto-correct the orientation of the pages
        auto_correct_orientation=False,
    )

    pipeline_options = PdfPipelineOptions(
        ocr_options=ocr_options,
    )
    pipeline_options.allow_external_plugins = True  # <-- enabled the external plugins

    # Convert the document
    converter = DocumentConverter(
        format_options={
            InputFormat.PDF: PdfFormatOption(
                pipeline_options=pipeline_options,
            ),
        },
    )

    conversion_result: ConversionResult = converter.convert(source=source)
    doc = conversion_result.document
    md = doc.export_to_markdown()
    print(md)


if __name__ == "__main__":
    main()

Configuration

The configuration of the OCR engine is done via the OnnxtrOcrOptions class. The following options are available:

• lang: List of languages to use for OCR. Default is ["en", "fr"].
• confidence_score: Word confidence threshold for the recognition model. Default is 0.5.
• objectness_score: Detection model objectness score threshold. Default is 0.3.
• det_arch: Detection model architecture. Default is "fast_base".
• reco_arch: Recognition model architecture. Default is "crnn_vgg16_bn".
• reco_bs: Batch size for the recognition model. Default is 512.
• auto_correct_orientation: Whether to auto-correct the orientation of the pages. Default is False.
• preserve_aspect_ratio: Whether to preserve the aspect ratio of the images. Default is True.
• symmetric_pad: Whether to use symmetric padding. Default is True.
• paragraph_break: Paragraph break threshold. Default is 0.035.
• load_in_8_bit: Whether to load the model in 8-bit. Default is False. (Not supported for Hugging Face loaded models yet)
• providers: List of providers to use for the Onnxruntime. Default is None which means auto-select.
• session_options: Session options for the Onnxruntime. Default is None which means default OnnxTR session options.

Available Hugging Face models can be found at Hugging Face.

Further information:

Please take a look at OnnxTR.

Contributing

Contributions are welcome!

Before opening a pull request, please ensure that your code passes the tests and adheres to the project's coding standards.

You can run the tests and checks using:

make style
make quality
make test

License

Distributed under the Apache 2.0 License. See LICENSE for more information.