Having inaccurate results for various Alpha Numeric Fields in English

[udayafo]

Help! I am getting low character accuracy for Images of English Documents.

I am using PP-OCRv4 with the following setup

ppocr_model_path = os.path.join(os.path.dirname(__file__), "model/en_PP-OCRv4_rec_infer")
ppocr_device = 'cpu'
ppocr_confidence = 0.25
ppocr_lang_type = "en"
model_instance = None

def init_ppocr_model(model_path, conf=0.5, lang="en"):
    global ppocr_model_path, ppocr_device, ppocr_confidence, ppocr_lang_type, model_instance

    ppocr_model_path = model_path
    ppocr_confidence = conf
    ppocr_lang_type = lang

    is_use_gpu = False
    if torch.cuda.is_available():
        ppocr_device = 'cuda'
        is_use_gpu = True

    model_ppocr = PaddleOCR(
        use_angle_cls=True,
        lang=ppocr_lang_type,
        use_gpu=is_use_gpu,
        rec_thresh=ppocr_confidence,
        rec_model_dir=ppocr_model_path
    )

    model_instance = model_ppocr

I have installed the following version of paddleocr:

(The font used is "Aptos" Bold if it helps)
Also we use unmodified OCR only and give it pre-segmented images. We also have some multi line OCR as well.

And examples of such would be:

• "I" being interpreted as "J", happens various times. Similarly "1", "7", "/" get interchanged with each other.
  • Input
    
  • Output
    
• Many times some characters and spaces are misread or missed entirely.
  • Input
    
  • Output
    
  • Input
    
  • Output
    

I need help if any possible through in regards to.

• What are some configuration options that may be tweaked to get better results.
• Are there any well known pre-processing techniques that may help in such scenarios.

Please share any and all ideas regarding this issue, thanks.

[liuhongen1234567]

Hello, you can try ppocrv5. Based on the current cases, it seems that ppocrv5 https://paddlepaddle.github.io/PaddleOCR/latest/en/version3.x/module_usage/text_recognition.html can solve them all.

[zhangyubo0722]

It is recommended to use latin_PP-OCRv5_mobile_rec, or wait for the paddleocr3.2 PP-OCRv5 English model.