Recommendation for docTR on CPU for Multilingual OCR

[mkrzywda]

Hi everyone,

I’m currently using docTR for OCR on GPU with the following configuration:

DET_ARCH = "fast_tiny"
REC_ARCH = "crnn_mobilenet_v3_small"
BIN_THRESH = 0.1
BOX_THRESH = 0.1    

model = ocr_predictor(det_arch=DET_ARCH, reco_arch=REC_ARCH, pretrained=True).to(device)                                           
doc = DocumentFile.from_images([tmp_path])
result = model(doc)

It works well, but I’m looking to switch to CPU and need recommendations for the best setup. My requirements are:

• Multilingual support: Mostly English, occasionally Polish.
• Low processing time: For images up to 4K resolution.
• Bounding box per word: High precision is crucial.

Any suggestions for optimal DET_ARCH, REC_ARCH, or other parameters to balance speed and accuracy on CPU?

Thanks!

[felixdittrich92]

Hi @mkrzywda 👋,

with all the requirements you should check out the following:

• OnnxTR latency benchmarks: https://github.com/felixdittrich92/OnnxTR#benchmarks
• multilingual (latin extended):
  • docTR: https://huggingface.co/Felix92/doctr-torch-parseq-multilingual-v1
  • OnnxTR: https://huggingface.co/Felix92/onnxtr-parseq-multilingual-v1

Hope this helps

Best,
Felix

[felixdittrich92]

Dataset I used to train this model can be downloaded here:
https://drive.google.com/file/d/1TNQN8uBMiGjzf2GM41BWDICefLubau5q/view?usp=sharing

Currently working on a better and more extended dataset

[mkrzywda]

What is a key-diff beetwen docTR and OnnxTR ? Because in docTR I can use more models than in OnnxTR ?

[felixdittrich92]

OnnxTR and docTR provides the same models(https://github.com/felixdittrich92/OnnxTR#text-detection), the key difference is that OnnxTR is a core refactoring of docTR which works with onnxruntime as backend instead of PyTorch or TensorFlow.
onnxruntime is hardware optimized, faster and more lightweight in general, additional it provides easy support for different accelerators - besides CPU / GPU (CUDA) also OpenVNO, DirectML, etc.

So the user interface between OnnxTR & docTR is mostly the same but the core differs - one difference: OnnxTR is inference only but optimized especially for prod systems and in docTR you have both inference & training logic

[mkrzywda]

Wonderful! I'm excited to try OnnxTR on CPU for processing images in this scenario. Is it possible to fine-tune models for my specific type of images?

[felixdittrich92]

Yes in this case use docTR to finetune the model/models to your own needs

1. https://mindee.github.io/doctr/using_doctr/custom_models_training.html
   then you can export the model to onnx
2. https://mindee.github.io/doctr/using_doctr/using_model_export.html#export-to-onnx
   and afterwards load in OnnxTR
3. https://github.com/felixdittrich92/OnnxTR#loading-custom-exported-models

:)