PaddleOCR/main/en/version3.x/module_usage/layout_detection

[giscus[bot]]

PaddleOCR/main/en/version3.x/module_usage/layout_detection

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://paddlepaddle.github.io/PaddleOCR/main/en/version3.x/module_usage/layout_detection.html

[AadelAbk-aiio]

When using Layout Detection, how can one extract the text inside each layout box, without having to use another model object and predict ? (when using a PDF)
For instance, I am using

model = LayoutDetection(model_name="PP-DocLayout_plus-L", layout_merge_bboxes_mode="large") # keep the largest outer box, remove inner overlapping boxes
layout_result = model.predict(pdf_path, batch_size=1, layout_nms=True)

To get the layout and it works fine. But I would like to different activity with different box labels identified.
Now I am using the bounding boxes, to crop the identified label as image and using PaddleOCR to Predict the text, but the results are not really nice as the images are tightly cropped and most of them are really small.
Is there way to add a parameter to the layout object so it has the details of the text from the respective labels, (if table have them also identified cell wise and give the result as list or pipe seperated text data, as other module library do??

[AadelAbk-aiio]

Unfortunately even after giving the option as False it still uses PP-Chart2Table.
I tried deleting the model and running it again, but it downloads it.

I also tried copying this code snippet and creating a new venv and starting from scratch, it downloads and uses the model even then. Could you please suggest what can be the issue ??

[liuhongen1234567]

Hello, PaddleOCR 3.0.1 should be able to solve this problem. This parameter in PaddleOCR 3.0.0 version will still load the chart model, but it will not call the chart model for prediction.

pip install paddleocr==3.0.1

[AadelAbk-aiio]

I tried that as well, it was giving me an error with mkldnn.

Exception has occurred: AttributeError
'paddle.base.libpaddle.AnalysisConfig' object has no attribute 'set_mkldnn_cache_capacity'
  File "/Users/aadel/Documents/GraphDB/Test1/test_paddlePipeline.py", line 12, in <module>
    pipeline = PPStructureV3(use_chart_recognition=False)
AttributeError: 'paddle.base.libpaddle.AnalysisConfig' object has no attribute 'set_mkldnn_cache_capacity'

However with paddleocr==3.0.0 like you said, it only loads the model doesn't use it seems.
I was able to generate an output for an image file, but not for a pdf. When attempted with PDf file, the terminal crashes (probably memory ran out).

[liuhongen1234567]

In version 3.0.0, if you disable the chart model, this error might still occur. It seems that your machine does not support MKLDNN. You can try disabling it by using enable_mkldnn=False.

from paddleocr import PPStructureV3
pipeline = PPStructureV3(use_chart_recognition=False, enable_mkldnn=False)
output = pipeline.predict("./pp_structure_v3_demo.png")
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") #

[liuhongen1234567]

If you only need OCR information, you can save memory by setting use_formula_recognition=False and use_table_recognition=False to further disable formula and table recognition.

[Atulok0506]

Hi,

I’m using the LayoutDetection module with the model PP-DocLayout_plus-L to perform layout detection on resumes. The problem I’m facing is that the model generates multiple bounding boxes that are very close to each other, especially in sections like Experience and Education.

For example, in the Experience section, the model creates one bounding box for every single line, but what I want is a single bounding box for the entire section.