PP-OCRv5 mobile (ONNX, for browsers)
ONNX versions of three PaddleOCR PP-OCRv5 models, for running in the browser with onnxruntime-web. They are used by the inference tab of IIIF Kasane.
| File | Original | What it does |
|---|---|---|
PP-OCRv5_mobile_det.onnx |
PaddlePaddle/PP-OCRv5_mobile_det (rev 0d63e78) | Finds text lines (DB) |
PP-OCRv5_mobile_rec.onnx |
PaddlePaddle/PP-OCRv5_mobile_rec (rev 682f205) | Reads a line: Simplified and Traditional Chinese, English, Japanese (CTC, 18,385 classes) |
latin_PP-OCRv5_mobile_rec.onnx |
PaddlePaddle/latin_PP-OCRv5_mobile_rec (rev ab2cd5c) | Reads a line in Latin-script languages (CTC, 838 classes) |
*.dict.json is the character_dict of each recognizer's inference.yml, as a JSON array. Class 0 is the CTC blank, class k + 1 is dict[k], and the last class is a space.
Conversion
The weights are unchanged (float32). Converted with paddle2onnx 2.1.0 (PaddlePaddle 3.3.1):
paddle2onnx --model_dir <model> --model_filename inference.json --params_filename inference.pdiparams --save_file <model>.onnx --opset_version 14
On random input, the ONNX outputs (onnxruntime 1.31, CPU) differ from Paddle inference by at most 1.3e-7 (detector) and 7.2e-5 (recognizer). On one printed English page (Internet Archive, Origin of Species, page 30, 1234 × 2048), a TypeScript port of the pipeline found the same 42 lines as PaddleOCR 3.7.0; 40 lines read identically, 3 of 2,266 characters differ.
License
Apache License 2.0, as the original models. Original models: PaddlePaddle Authors. Conversion: Satoru Nakamura (The University of Tokyo).