latest/en/version3.x/pipeline_usage/PP-StructureV3

[giscus[bot]]

latest/en/version3.x/pipeline_usage/PP-StructureV3

Awesome multilingual OCR toolkits based on PaddlePaddle (practical ultra lightweight OCR system, support 80+ languages recognition, provide data annotation and synthesis tools, support training and deployment among server, mobile, embedded and IoT devices)

https://www.paddleocr.ai/latest/en/version3.x/pipeline_usage/PP-StructureV3.html

[Lafitte1573]

here is my script:

from paddleocr import PPStructureV3

input_file = "3ul2Rq4Sk5Cn-l69D695U.png"
pipeline = PPStructureV3(
    # models
    text_recognition_model_name="PP-OCRv5_mobile_rec",
    text_detection_model_name="PP-OCRv5_mobile_det",
    # table_classification_model_name="PP-LCNet_x1_0_table_cls",
    # wired_table_cells_detection_model_name="SLANeXt_wired",
    formula_recognition_model_name="PP-OCRv5_mobile_rec",  # "PP-OCRv5_mobile_rec",
    layout_detection_model_name="PP-DocBlockLayout",  # "PP-DocBlockLayout",
    region_detection_model_name="PP-DocBlockLayout",
    # model dirs
    text_recognition_model_dir='/mnt/public/public_file/models/PP-OCRv5_mobile_rec',
    text_detection_model_dir='/mnt/public/public_file/models/PP-OCRv5_mobile_det',
    # table_classification_model_dir='/mnt/public/public_file/models/PP-LCNet_x1_0_table_cls',
    # wired_table_cells_detection_model_dir="/mnt/public/public_file/models/SLANeXt_wired",
    formula_recognition_model_dir='/mnt/public/public_file/models/PP-OCRv5_mobile_rec',
    layout_detection_model_dir='/mnt/public/public_file/models/PP-DocBlockLayout',
    region_detection_model_dir='/mnt/public/public_file/models/PP-DocBlockLayout',
    # Use options
    use_doc_orientation_classify=False,
    # # Use use_doc_orientation_classify to enable/disable document orientation classification model
    use_doc_unwarping=False,  # Use use_doc_unwarping to enable/disable document unwarping module
    use_textline_orientation=False,
    # # Use use_textline_orientation to enable/disable textline orientation classification model
    use_seal_recognition=False,
    use_table_recognition=False,
    use_chart_recognition=False,
    device="gpu",  # Use device to specify GPU for model inference
)
output = pipeline.predict(input=input_file)

for res in output:
    res.print() # Print the structured prediction output
    res.save_to_json(save_path="output") ## Save the current image's structured result in JSON format
    res.save_to_markdown(save_path="output")

Why is my output Markdown file blank?

[liuhongen1234567]

Hello, you can try setting the layout_detection_model_dir to PP-DocLayout_plus-L. PP-DocBlockLayout is a block detection model that only detects the block category, not a layout model.

[Lafitte1573]

Thanks a lot. I changed the layout_detection_model to PP-DocLayout_plus-L, and this problem was successfully fixed. 😄😄😄

[w2918000]

请问下面dict的中类别ID是在哪里定义的？

layout_threshold 版面模型得分阈值。
float：0-1 之间的任意浮点数；
dict： {0:0.1} key为类别ID，value为该类别的阈值；
None：如果设置为None，将使用产线初始化的该参数值，默认初始化为 0.5。

[w2918000]

或者说每一项id具体对应的目标是什么

[liuhongen1234567]

您好，可以下载具体的版面模型然后查看inference.yaml 的 label_list 文件，从0开始依次编号。 如果是 PP-Structurev3的默认版面模型可以直接参考 PP-Structurev3 产线配置文件

https://github.com/PaddlePaddle/PaddleX/blob/1f63f83829c8242a1655127b760206ae5faa30e8/paddlex/configs/pipelines/PP-StructureV3.yaml#L19

[naked34501]

你好 ，我想问下 版面识别是否可以在识别的时候自动忽略印章？检验合同文档OCR的时候如何解决文字被印章覆盖的问题，导致OCR出来包含部分印章文字？同时印章的文字识别也被原本的合同文字遮挡导致失败？是否能进行两边的分离检查？万分感谢！

[w2918000]

你好，有个问题咨询下，我想知道在使用PP-StructureV时如何调整表格识别下单元格识别的threshold参数

[hykyle]

公式乱码：
$$P_{\mathrm{SRD}}=\frac{3.6nK_{\mathrm{DL}}K_{\mathrm{DY}}}{tK}\times10^3\quad\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots P P_{P_{{{\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots P_{{\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots P_\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots P_\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots P_\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots P_\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots P_{\cdots\cdots\cdots\cdots\cdots P_{\cdots\cdots\cdots\cdots\cdots\cdots P_\cdots\cdots\cdots\cdots\cdots P_{\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots\cdots P_\cdots\cdots\cdots\cdots\cdots\cdots\cdots P_\cdots\cdots\cdots\cdots P_{\cdots\cdots\cdots\cdots\cdots\cdots P_\cdots\cdots\cdots\cdots\cdots\cdots P_\cdots\cdots\cdots\cdots P_{\cdots\cdots\cdots\cdots\cdots P_\cdots\cdots\cdots\cdots P_\cdots\cdots\cdots\cdots\cdots P_\cdots\cdots\cdots P_\cdots{\cdots\cdots P_\cdots\cdots\cdots P_\cdots\cdots\cdots\cdots P_\cdots\cdots\cdots\cdots P_\cdots\cdots\cdots P_{\cdots\cdots\cdots\cdots P_\cdots\cdots\cdots P_\cdots\cdots\cdots\cdots P_\cdots\cdots P_\cdots P_{\cdots\cdots\cdots\cdots\cdots P_\cdots\cdots\cdots\cdots P_\cdots P_\cdots\cdots\cdots\cdots\cdots P_\cdots\cdots P_\cdots\cdots P_{\cdots\cdots\cdots P_\cdots\cdots P_\cdots\cdots\cdots P_\cdots\cdots\cdots P_\cdots\cdots P_\cdots\cdots P_\cdots\cdots P_\cdots\cdots\cdots P_\cdots\cdots P_\cdots P_{\cdots\cdots P_\cdots P_\cdots\cdots\cdots\cdots P_\cdots P_\cdots\cdots\cdots P_\cdots\cdots P_\cdots P_\cdots\cdots P_\cdots\cdots P_\cdots P_{\cdots\cdots\cdots\cdots P_\cdots P_\cdots\cdots P_\cdots P_\cdots P_\cdots P_\cdots\cdots P_\cdots P_{\cdots P_\cdots\cdots P_\cdots\cdots\cdots P_\cdots P_\cdots\cdots P_\cdots P_\cdots P_\cdots P_\cdots P_\cdots\cdots P_\cdots P_\cdots P_\cdots P_\cdots P_{\cdots P_\cdots P_\cdots P_\cdots\cdots P_\cdots P_\cdots\cdots P_\cdots P_{\cdots P_\cdots P_\cdots P_\cdots P_\cdots P_\cdots P_\cdots\cdots\cdots P_\cdots P_\cdots P_\cdots P_\cdots P_\cdots\cdots P_\cdots P_\cdots\cdots P_\cdots P_\cdots P_\cdots\cdots P_\cdots P_\cdots P_{\cdots P_\cdots P_\cdots P_\cdots P_\cdots\cdots P_\cdots P_\cdots P_\cdots P_\cdots P_\cdots\cdots P_\cdots P_\cdots P_\cdots P_\cdots\cdots P_\cdots P_\cdots P_\cdots{\cdots P_\cdots P_\cdots P_\cdots\cdots P_\cdots P_\cdots P_\cdots P_\cdots $$

[liuhongen1234567]

您好，能否提供一下测试的图片呢？由于公式采用的是自回归架构，在一些域外数据上确实可能出现幻觉，导致输出乱码。

[hykyle]

《070 泵类系统电能平衡测试与计算方法-GBT13468-1992.pdf》这篇国标文档。找不到上传图片或文件的地方

[hykyle]

不对表格做提取，但如果表格中有公式，则表格图片中的公式就会被抠掉。咋整？？

[Mirorrn]

Hi :)
Is there an official way to extract tables with captions?

[deepak7376]

can we quantize the model into less precision like int8 for this pipeline

from paddleocr import PPStructureV3

[loooau1]

你好可以问下这个模型可以并发使用吗？

[Echo-Forever]

你好 ，我想问下 版面识别是否可以在识别的时候自动忽略印章？检验合同文档OCR的时候如何解决文字被印章覆盖的问题，导致OCR出来包含部分印章文字？同时印章的文字识别也被原本的合同文字遮挡导致失败？是否能进行两边的分离检查？万分感谢！

[BNN0]

I'm a little lost. How can I use paddleocr PPStructureV3 completely locally from a specific directory with the models?
This is because when I have the following configuration, it continues to download files to the original paddleocr path, and I need it to only use the models hosted in a directory within my project.

from paddleocr import PPStructureV3

pipeline = PPStructureV3(
text_recognition_model_name="PP-OCRv5_server_rec",
text_detection_model_name="PP-OCRv5_server_det",
formula_recognition_model_name="PP-FormulaNet_plus-L",
layout_detection_model_name="PP-DocLayout_plus-L",
region_detection_model_name="PP-DocBlockLayout",

text_recognition_model_dir='./models/PP-OCRv5_server_rec',
text_detection_model_dir='./models/PP-OCRv5_server_det',
formula_recognition_model_dir='./models/PP-FormulaNet_plus-L',
layout_detection_model_dir='./models/PP-DocLayout_plus-L',
region_detection_model_dir='./models/PP-DocBlockLayout',

use_doc_orientation_classify=False,
use_doc_unwarping=False,
use_seal_recognition=False,
use_table_recognition=False,
use_chart_recognition=False,

)

"models" is my local folder for dowloaded models for PPStructureV3

And how can I disable that download validation that is performed every time the code is executed? It's annoying to have to wait several seconds for the files to be validated before I can even start the process.

[javierjdaza]

im having this error

local variable 'transpose_weight_keys' referenced before assignment

My code:

from paddleocr import PPStructureV3

pipeline = PPStructureV3(
    lang='pt',
    use_doc_orientation_classify=True,
    use_doc_unwarping=True,
    use_textline_orientation=True,
)

Any Help?

[ivanart555]

I'm having a similar problem.

packages\paddlex\inference\models\doc_vlm\predictor.py", line 131, in _build model = PPChart2TableInference.from_pretrained( ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "G:\Python\Lib\site-packages\paddlex\inference\models\common\vlm\transformers\model_utils.py", line 1940, in from_pretrained transpose_weight_keys=transpose_weight_keys, ^^^^^^^^^^^^^^^^^^^^^ UnboundLocalError: cannot access local variable 'transpose_weight_keys' where it is not associated with a value

[Surya0416]

All are good table detection text detection all are good but the html construction is not good why it happens?

[nguyendola]

I want to ask when Vietnamese text will be supported.

[iris-qq]

为什么没有一个office_model得参数设置，因为有七个模型，我要写7次地址，另外这个默认模型的下载地址也没有给出具体参数来配置路径，很麻烦，全部给我下载到c盘了

[iris-qq]

我有个使用问题，就是输出的markdown文本，如果包含图片是如下格式：

<div style="text-align: center;"><img src="https://Fupload%2F20251219%2F9a97a0cb01b644f981ab60eb50788b71.png" alt="Image" width="63%" /></div>


取料机构中的滚筒表面要保持洁净无凹凸不平现象，每天可用酒精清洗，若需要跟换，可用5MM六角扳手拆掉（如下图所示a处)螺丝，拆掉(b处）侧板，即可跟换。



<div style="text-align: center;"><img src="https://F01d540938c424f18b83f55b5d32612fc.png" alt="Image" width="70%" /></div>

但实际上markdown的样式是 !(测试图片)[url]这种，请问有什么参数去解决这个问题么

[yonglee7015]

我怎么感觉你们的代码老是这不兼容，那不兼容的，你们维护的代码简直像一坨翔？就不会安排人整理下吗？ 直接按你们的要求安装各种库，都无法完美的跑你们的处理pdf的示例代码。。。。。虽然paddleocr我觉得很牛逼。
PP-StructureV3跑你的示例代码

File "D:\VSstudioProjects\paddleocr_gpu.venv\Lib\site-packages\paddlex\inference\pipelines\layout_parsing\pipeline_v2.py", line 496, in standardized_data
layout_det_res["boxes"].append(
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
AttributeError: 'numpy.ndarray' object has no attribute 'append'

[code-frog-one]

如何设置 return_word_box参数未True，我在 GeneralOCR:
pipeline_name: OCR
batch_size: 8
text_type: general
use_doc_preprocessor: False
use_textline_orientation: True
return_word_box: True
SubModules:
TextDetection:
module_name: text_detection
model_name: PP-OCRv5_server_det
model_dir: null
limit_side_len: 736
limit_type: min
max_side_limit: 4000
thresh: 0.3
box_thresh: 0.6
unclip_ratio: 1.5
TextLineOrientation:
module_name: textline_orientation
model_name: PP-LCNet_x1_0_textline_ori
model_dir: null
batch_size: 8
TextRecognition:
module_name: text_recognition
model_name: PP-OCRv5_server_rec
model_dir: null
batch_size: 8
score_thresh: 0.0
中设置了没有生效