Alphabet for English rec model is not matching OCR on image

[christegho]

Thank you so much for providing code to convert paddle models to onnx format and be able to use these converted models with rapid OCR.

I used paddleocr_convert to convert the latest english models as follow:

paddleocr_convert -p  https://paddleocr.bj.bcebos.com/PP-OCRv4/english/en_PP-OCRv4_rec_infer.tar -o paddle_models  -txt_path https://raw.githubusercontent.com/PaddlePaddle/PaddleOCR/release/2.7/ppocr/utils/en_dict.txt

paddleocr_convert -p  https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_det_infer.tar  -o paddle_models

For testing these models, I use the following image

I use rapid ocr to get OCR:

from rapidocr_onnxruntime import RapidOCR

engine = RapidOCR(rec_model_path="paddle_models/en_PP-OCRv4_rec_infer/en_PP-OCRv4_rec_infer.onnx", 
det_model_path="paddle_models/en_PP-OCRv3_det_infer/en_PP-OCRv3_det_infer.onnx")

img = "c8fa66439d63041e.jpg"
result, elapse = engine(img)

and I get the following result (I only included the first few predictions):

[[[[551.0, 72.0], [952.0, 83.0], [949.0, 177.0], [548.0, 166.0]],
  'DlizAly',
  0.8711231350898743],
 [[[540.0, 211.0], [978.0, 221.0], [978.0, 250.0], [540.0, 241.0]],
  'ulD ADluAplAo jluABy',
  0.9064718882242838],
 [[[65.0, 280.0], [207.0, 277.0], [207.0, 293.0], [65.0, 297.0]],
  "B!QR%%'R}N}   T",
  0.6442324388772249],
 [[[220.0, 286.0], [270.0, 283.0], [270.0, 297.0], [220.0, 300.0]],
  'B!QV&P"%R',
  0.7667727172374725],
...

Whereas the default models produce more matching results

from rapidocr_onnxruntime import RapidOCR

engine = RapidOCR()

img = "c8fa66439d63041e.jpg"
result, elapse = engine(img)

result -->

[[[[545.0, 72.0], [954.0, 84.0], [951.0, 179.0], [542.0, 167.0]],
  'WEBSTER',
  0.8526573106646538],
 [[[539.0, 210.0], [975.0, 218.0], [975.0, 250.0], [539.0, 243.0]],
  'NEW TWENTIETH CENTUR',
  0.9136510831969125],
 [[[63.0, 280.0], [209.0, 277.0], [209.0, 295.0], [63.0, 299.0]],
  'Underrte, r.aa ferg',
  0.6786298751831055],
 [[[220.0, 285.0], [271.0, 282.0], [271.0, 298.0], [220.0, 301.0]],
  'Undlacove',
  0.7574679881334305],

This suggests I am perhaps using the wrong dictionary. However, the only dict I could find in the paddle git repo is https://raw.githubusercontent.com/PaddlePaddle/PaddleOCR/release/2.7/ppocr/utils/EN_symbol_dict.txt.
Do you happen to know if there is some step I am doing wrong here?

thank you so much,

Chris

[christegho]

I am closing this issue. After posting the issue, I have noticed that I was using a local dict that was slightly different from the one available with paddle ocr