Arabic: list index out of range

[matteoPiccoli]

Hello everyone!

My testing dataset is composed of nearly 60 images, and I'm having troubles analyzing the following two images:

I think that the problem could be the recognition of these characters:

1. :
2. '

More in details, text detection works perfectly but when it comes the recognition stage the following error shows up:

Traceback (most recent call last):
  File "c:\Users\matte\OneDrive\Documenti\GitHub\ATB-Text-Recognition\atbocr\engine\predictions.py", line 44, in paddle_recognition
    result = Paddle.ocr(cropped_image, cls=False, det=False, rec=True)
  File "C:\Python\Python39\lib\site-packages\paddleocr\paddleocr.py", line 538, in ocr
    rec_res, elapse = self.text_recognizer(img)
  File "C:\Python\Python39\lib\site-packages\paddleocr\tools\infer\predict_rec.py", line 532, in __call__
    rec_result = self.postprocess_op(preds)
  File "C:\Python\Python39\lib\site-packages\paddleocr\ppocr\postprocess\rec_postprocess.py", line 120, in __call__
    text = self.decode(preds_idx, preds_prob, is_remove_duplicate=True)
  File "C:\Python\Python39\lib\site-packages\paddleocr\ppocr\postprocess\rec_postprocess.py", line 82, in decode
    char_list = [
  File "C:\Python\Python39\lib\site-packages\paddleocr\ppocr\postprocess\rec_postprocess.py", line 83, in <listcomp>
    self.character[text_id]
IndexError: list index out of range

The Paddle model is initialized as follows:

PaddleOCR(use_angle_cls=True, lang="ar", use_gpu=True)

Thank you

[tink2123]

1. The : ' character cannot be recognized because there is a lack of similar data in the training data. If necessary, you can refer to the training document and use your own data finetune.

2. What version of paddleocr is installed? You can try to update to paddleocr==2.6.0.1 . The new version of the whl package should fix the IndexError.

[matteoPiccoli]

I already installed the 2.6.0.1 version of paddleocr

[tink2123]

Can you provide detailed output information for prediction? Let me check whether the model and the dictionary are consistent. Or you can update the Arabic model:

wget https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/arabic_PP-OCRv3_rec_infer.tar 

tar -xf arabic_PP-OCRv3_rec_infer.tar

Modify the recognition model path:

PaddleOCR(use_angle_cls=True, lang="ar", use_gpu=True, rec_model_dir="arabic_PP-OCRv3_rec_infer")