pdfdet: PDF Layout Detection Toolbox

Requirements

python>=3.10

python -m pip install -r requirements.txt
pip install .

Usage

Simple Operations

python main.py --path "pdf_path"

You can also import a model directly via its full name and then call its __call__ method with pdf path or image path.

• Layer

  from pdfdet import uni_model
  model = uni_model(name='yolov8m_cdla')
  layer = model(path="image_path")
  content  = layer.to_json()
  """
  content format:
  {
      "image": numpy.ndarray,
      "boxes": [{"box": [x1, y1, x2, y2], "label": str, "score": float}, ...],
  }
  """

• Document

  from pdfdet import uni_model
  import cv2
  model = uni_model(name='yolov8m_cdla')
  doc = model(path="pdf_path")
  layers = sorted(doc.layers, key=lambda x: int(x))
  for i in layers:
      layer = getattr(doc, i)
      im = layer.imshow()
      im = cv2.resize(im, (640, 640))
      cv2.imshow("im", im)
      cv2.waitKey(0)
  cv2.destroyAllWindows()

Batch processing

# batch predict
python tools/batch_process.py --model "model_name" --src "image_root" --save "res_root"
# generate visualize result
python tools/visualize.py "image_path" "res_path"
#
# evaluate cdla dataset(labelme format)
python tools/eval_map50.py "gt_root" "res_root"

Models

Model	Source	Associated Dataset	optional model
paddle_pub	PaddlePaddle	PubLayNet(English)
paddle_cdla	PaddlePaddle	CDLA(Chinese)
cnstd_yolov7	CNSTD	CDLA
yolov8l_doc	huggingface	DocLayNet(English, German,French, Japanese)	yolov8n_doc,yolov8s_doc
yolov8m_cdla	layout_analysis	CDLA	yolov8n_cdla

Evaluation Code Source

Note: Labels and annotation strategies vary across different datasets. Visual comparison should be the primary method for evaluating effectiveness.

CDLA

CDLA

Model	map50	map50:95	p	r
paddle_cdla	0.9675	0.8359	0.9602	0.9347
cnstd_yolov7	0.9058	0.6662	0.9543	0.8321
yolov8m_cdla	0.9436	0.8086	0.9449	0.8980

mnbvc

Test Dataset

Model	map50:95	p	r
paddle_cdla	0.5717	0.5853	0.6248
cnstd_yolov7	0.5034	0.6278	0.5651
yolov8m_cdla	0.4783	0.5266	0.5922