arabic recognition model gives bad results for arabic numbers

[amroghoneim]

I am going through the tutorial and testing out some images that include only arabic numbers.

I usually see bad results even though its quite accurate for text.

Can this be resolved using some parameters? or do I need to finetune the recognition model ?

[mohamedbakhet]

I face same problem

[amroghoneim]

@MissPenguin Please advise.

[LDOUBLEV]

Try finetune training. (The data we have is really limited.)

If it is a bad case, it is recommended to send some samples to analyze.

[Mahmuod1]

@LDOUBLEV can you provide a sample of the Arabic dataset you used in the training

[tink2123]

@amroghoneim Could you provide prediction commands and badcase? Let's reproduce the effect.

[amroghoneim]

Sure @tink2123

So this is a sample code

from paddleocr import PaddleOCR,draw_ocr


ocr = PaddleOCR( lang='ar', max_text_length = 50, rec_batch_num = 1) # need to run only once to download and load model into memory
result = ocr.ocr(np.array(img), cls=False, det = False)

example images with numbers only:

By the way I also see degradation when combining numbers with text, and when doing recognition on multiple words compared to splitting them each one on their own.

[Topdu]

Thanks! This problem is caused by incorrectly processing the Arabic number labels during training, and we will subsequently update the model after retraining. You can also perform finetune training on the Arabic numbers dataset to solve this problem.

[amroghoneim]

@Topdu will this be released anytime soon? as we would really like to use this great OCR solution in our codebase.

Also the finetuning process seems to be hard for me to get my head around when following the documentation. May I ask here what are the steps I need to follow exactly? how should the dataset look like (example dataset url is not working)? is there a training script I can use off the bat? What does the GT label file look like and so on.

[amroghoneim]

@Topdu Any comments please?

[Topdu]

As soon as possible.

For training documents, please refer to rec training doc.

[AstonyJ]

遇到了同样的问题，使用latin_PP-OCRv3_rec模型，想问下是否有更新之后的模型呢？

[vk1996]

The arabic numbers prediction of SVTR arabic model is poor . any updates on this please ?

[Topdu]

PP-OCRv4 arabic model

[vk1996]

is the config file same as v3 ?

[Hegelim]

PP-OCRv4 arabic model

This is an inference model, what about pre-trained model for Arabic?

[Hegelim]

Is the dictionary same?