sbb-binarization-onnx

SBB Binarization — ONNX Model

ONNX conversion of SBB/sbb_binarization by the Berlin State Library (Staatsbibliothek zu Berlin), developed as part of the QURATOR project.

The original model is a UNet + Vision Transformer hybrid that converts scanned document images to black and white for OCR. It works on 448x448 patches.

What's here

• model_convtranspose.onnx — the model, ready to use with ONNX Runtime
• fix_onnx.py — the script that converts SBB's TensorFlow SavedModel to this ONNX file
• sample_workflow.py — minimal example showing how to binarize an image
• example_gpu_pipeline.py — full GPU pipeline using CuPy for production use

Quick start

pip install onnxruntime-gpu numpy Pillow
python3 sample_workflow.py input.jpg output.tif

What was changed from the original

The original TF model doesn't convert cleanly to ONNX for TensorRT. Three things needed fixing:

1. Reshape batch dimension — tf2onnx hardcoded -2048 instead of -1 for dynamic batching
2. Resize node attributes — TF-specific modes that TensorRT doesn't support, swapped to standard ONNX equivalents
3. Resize to ConvTranspose — replaced nearest-neighbor upsampling with equivalent depthwise ConvTranspose ops so TensorRT can compile the model as a single subgraph instead of splitting it into 8+ pieces

All of this is in fix_onnx.py. To reproduce from scratch:

pip install tf2onnx onnx tensorflow
python3 -m tf2onnx.convert --saved-model path/to/saved_model/2022-08-16 --output model.onnx --opset 17
python3 fix_onnx.py model.onnx model_convtranspose.onnx

Output is <0.01% pixel difference from the original TF model.

License

Same as the original — Apache 2.0. All credit to the SBB team and the QURATOR project for the model itself.