Trained model ppocrv5_server_rec does not recognize text

[KhankhaevAndrey]

Hi! Trained the ppocrv5_server_rec model on my handwritten numbers dataset. After training, it does not recognize the same images at all. I understand that it could be a training problem. But maybe there are more obvious reasons? images of size (79, 176, 3). example

I will attach the inference config, code and recognition results.
inference config.yml
Global:
model_name: PP-OCRv5_server_rec
Hpi:
backend_configs:
paddle_infer:
trt_dynamic_shapes: &id001
x:
- - 1
- 3
- 48
- 160
- - 1
- 3
- 48
- 320
- - 8
- 3
- 48
- 3200
tensorrt:
dynamic_shapes: *id001
PreProcess:
transform_ops:

• DecodeImage:
  channel_first: false
  img_mode: BGR
• MultiLabelEncode:
  gtc_encode: NRTRLabelEncode
• RecResizeImg:
  image_shape:
  • 3
  • 48
  • 320
• KeepKeys:
  keep_keys:
  • image
  • label_ctc
  • label_gtc
  • length
  • valid_ratio
    PostProcess:
    name: CTCLabelDecode
    character_dict:
• '0'
• '1'
• '2'
• '3'
• '4'
• '5'
• '6'
• '7'
• '8'
• '9'
• ','
• .

code

ocr = PaddleOCR(text_recognition_model_name="PP-OCRv5_server_rec",text_recognition_model_dir=r"C:\Users\Andrew\anaconda3\envs\v5train\Lib\site-packages\PaddleOCR\train_modelv5\inference",
                use_doc_orientation_classify=False,
                use_textline_orientation=False,
                use_doc_unwarping=False,
                lang='en')

result = ocr.predict(img_path)
print(result)

result

[{'input_path': 'C:\\Users\\Andrew\\Jupyterprojects\\Diploma\\cells\\cell_1_6_image_6_2.jpg', 'page_index': None, 'doc_preprocessor_res': {'input_path': None, 'page_index': None, 'input_img': array([[[255, ..., 255],
        ...,
        [255, ..., 255]],

       ...,

       [[255, ..., 255],
        ...,
        [255, ..., 255]]], dtype=uint8), 'model_settings': {'use_doc_orientation_classify': True, 'use_doc_unwarping': True}, 'angle': 180, 'rot_img': array([[[  0, ...,   0],
        ...,
        [  0, ...,   0]],

       ...,

       [[  0, ...,   0],
        ...,
        [255, ..., 255]]], dtype=uint8), 'output_img': array([[[255, ..., 255],
        ...,
        [255, ..., 255]],

       ...,

       [[254, ..., 254],
        ...,
        [255, ..., 255]]], dtype=uint8)}, 'dt_polys': [array([[47, 29],
       ...,
       [47, 51]], dtype=int16), array([[63, 35],
       ...,
       [63, 36]], dtype=int16), array([[63, 35],
       ...,
       [63, 35]], dtype=int16), array([[65, 35],
       ...,
       [65, 36]], dtype=int16), array([[65, 35],
       ...,
       [65, 36]], dtype=int16), array([[66, 36],
       ...,
       [66, 36]], dtype=int16), array([[66, 44],
       ...,
       [66, 45]], dtype=int16), array([[67, 38],
       ...,
       [67, 38]], dtype=int16), array([[43, 48],
       ...,
       [45, 51]], dtype=int16), array([[83, 42],
       ...,
       [83, 43]], dtype=int16), array([[85, 41],
       ...,
       [85, 42]], dtype=int16), array([[85, 42],
       ...,
       [85, 43]], dtype=int16), array([[100,  45],
       ...,
       [100,  46]], dtype=int16)], 'model_settings': {'use_doc_preprocessor': True, 'use_textline_orientation': False}, 'text_det_params': {'limit_side_len': 736, 'limit_type': 'min', 'thresh': 0.3, 'max_side_limit': 4000, 'box_thresh': 0.6, 'unclip_ratio': 1.5}, 'text_type': 'general', 'text_rec_score_thresh': 0.0, 'rec_texts': ['', '', '', '', '', '', '', '', '', '', '', '', ''], 'rec_scores': [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], 'rec_polys': [array([[47, 29],
       ...,
       [47, 51]], dtype=int16), array([[63, 35],
       ...,
       [63, 36]], dtype=int16), array([[63, 35],
       ...,
       [63, 35]], dtype=int16), array([[65, 35],
       ...,
       [65, 36]], dtype=int16), array([[65, 35],
       ...,
       [65, 36]], dtype=int16), array([[66, 36],
       ...,
       [66, 36]], dtype=int16), array([[66, 44],
       ...,
       [66, 45]], dtype=int16), array([[67, 38],
       ...,
       [67, 38]], dtype=int16), array([[43, 48],
       ...,
       [45, 51]], dtype=int16), array([[83, 42],
       ...,
       [83, 43]], dtype=int16), array([[85, 41],
       ...,
       [85, 42]], dtype=int16), array([[85, 42],
       ...,
       [85, 43]], dtype=int16), array([[100,  45],
       ...,
       [100,  46]], dtype=int16)], 'textline_orientation_angles': [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], 'rec_boxes': array([[ 47, ...,  51],
       ...,
       [100, ...,  46]], dtype=int16)}]

Thank you for answer

[KhankhaevAndrey]

I learned that if you resize the image by increasing the white background, the recognition results improve. But it is very unstable, if you change the size of the image by a couple of pixels, it will not be recognized.

this image is recognized, but if I resize it a bit, it will not be recognized.
Each other image must be resized individually as well.

[liuhongen1234567]

Hello, the PaddleOCR method uses the OCR pipeline, which first performs detection and then recognition. It seems that the issue might be with the detection part. If you want to recognize text from an already cropped image, it might be better to use the text recognition module https://paddlepaddle.github.io/PaddleOCR/main/en/version3.x/module_usage/text_recognition.html directly. Here is a sample code:

from paddleocr import TextRecognition
model = TextRecognition(model_name="PP-OCRv5_server_rec", model_dir=r"C:\Users\Andrew\anaconda3\envs\v5train\Lib\site-packages\PaddleOCR\train_modelv5\inference")
output = model.predict(input=img_path, batch_size=1)
for res in output:
    res.print()
    res.save_to_img(save_path="./output/")
    res.save_to_json(save_path="./output/res.json")

[liuhongen1234567]

When training the detection model, larger document images are used as input. If you input a cropped area that is meant to be recognized, it might indeed not be detected.

[KhankhaevAndrey]

When training the detection model, larger document images are used as input. If you input a cropped area that is meant to be recognized, it might indeed not be detected.

I'm very grateful to you! It was very important for me. PaddleOCR integration and training was needed for my graduation project. You are very cool. Thank you!!!