Need help with PP-OCRv5 onnx with a custom dictionary

[vqgquan]

So i have successfully export my model to onnx but when i run it this is what it return 1A5A9A41A5A1A171A6A5A while the correct prediction would be 048304006054. Is there something that im doing wrong?

[liuhongen1234567]

Hello, could you provide the reproduction code? Additionally, you can try using the official paddle2onnx method. The documentation is as follows: https://www.paddleocr.ai/latest/version3.x/deployment/obtaining_onnx_models.html

[vqgquan]

onnx_ocr.ipynb
this is my current notebook for testing ppocrv5 onnx. Is my decoding wrong?
inference.yml
This is my inference.yml created using paddle2onnx

[vqgquan]

I also found that the onnx model expect 212 classes while my dictionary from training only have 210 classes. What are the 2 classes that i'm missing?

[liuhongen1234567]

Hello, it seems that there is an issue with your decoding process. In PaddleOCR, besides the dictionary, there are some other special characters involved in the decoding. You can make modifications based on the code https://github.com/PaddlePaddle/PaddleX/blob/9579f20886f80b795bb75ad48097b3bebae744d3/paddlex/inference/models/text_recognition/processors.py#L269 here.

[vqgquan]

I have add at the beginning of the my dictionary file and at the end of it. This have make the prediction in the correct structurally but it is still not mapping correctly. It now output ỵ372ỵ3ỵỵ5ỵ43 instead of 048304006054.
Here is my current dictionary file:
vi_vietnamese.txt
This is my updated notebook:
onnx_ocr.ipynb