PDF to Markdown Converter (with YOLO-based Layout Detection)

This project provides a pipeline to convert academic/research PDFs into cleaned Markdown files, with support for:

• Trimming margins and removing line numbers.
• YOLO-based region detection (text, tables, figures, formulas).
• Post-processing filters for noise, licenses, and references.
• Markdown export with embedded images for figures/tables.
• Optional raw JSONL outputs for downstream processing.

---

📂 Project Structure

.
├── main.py
├── pdf_extractor.py
├── markdown_coverter.py
├── yolo_model/
│   ├── YoloModel.py
│   ├── YoloHelper.py
│   ├── YoloPipline.py
│   └── doclaynet.pt   # YOLO model weights (DocLayNet)
├── pdf_processor/
│   ├── NumberPaper.py
│   └── PdfTrimmer.py
├── text_filters/
│   ├── LicenseFilter.py
│   └── ReferenceFilter.py

---

⚙️ Modules Overview

1. PDF Preprocessing

• PdfTrimmer.py – trims page margins (top, bottom, left, right) while avoiding rotated pages.
• NumberPaper.py – detects and removes line numbers from PDF margins using heuristics.

2. YOLO-based Layout Extraction

• YoloModel.py – runs YOLO (DocLayNet weights) on PDF pages to detect text, figures, tables, and formulas.
• YoloHelper.py – processes YOLO outputs: merges bounding boxes, removes noise, raises headers, and restructures text.
• YoloPipline.py – orchestrates detection + post-processing into JSONL outputs.

3. Text Processing & Export

• markdown_coverter.py – converts processed JSONL into Markdown format, embedding images when available.
• LicenseFilter.py – removes boilerplate license/rights text.
• ReferenceFilter.py – removes references/bibliographies and cleans extraneous text.

4. Pipeline Entry Points

• pdf_extractor.py – core pipeline that:
  1. Trims PDFs and removes line numbers.
  2. Runs YOLO layout extraction.
  3. Applies license/reference filters.
  4. Saves .md, .jsonl, and removed sections.
• main.py – example script to run extraction on a folder of PDFs.

---

🚀 Usage

1. Install Dependencies

pip install -r requirements.txt

Dependencies include:

• PyMuPDF (fitz)
• PyPDF2
• ultralytics (YOLOv8)
• Pillow
• tqdm

2. Prepare Model Weights

The first time you run the pipeline, weights_utils.py will automatically download the DocLayNet YOLOv8X checkpoint from Hugging Face (malaysia-ai/YOLOv8X-DocLayNet-Full-1024-42) and save it as: yolo_model/doclaynet.pt

On subsequent runs, the script will detect the existing file and reuse it, so you don’t need to manually manage the weights.

3. Run the Pipeline

python main.py <input_folder> <output_folder> [--save-raw-json] [--save-removed]

Arguments

• input_folder – Path to the folder containing PDFs (e.g., paper/前列腺癌)
• output_folder – Name of the dataset/output folder (e.g., 前列腺癌).
• --save-raw-json – (optional) Save raw JSONL outputs from YOLO post-processing.
• --save-removed – (optional) Save removed license/reference sections.

4. Outputs

After running, you’ll get:

output_folder/
│── outputs/
│   ├── <paper_name>.md          # Final Markdown
│── outputs/images/              # Extracted figures/tables
│── raw_outputs/                 # JSONL structured outputs
│── removed/                     # Removed license/reference sections

---

🧩 Example Workflow

1. Place your PDFs into paper/<topic>/.
2. Run python main.py paper/<topic>/ outputs.
3. Check outputs/ for Markdown files.
4. Open .md files in any Markdown editor to view cleaned text + figures.

---

📜 License

This project is provided for research and educational purposes.
Make sure you comply with dataset (DocLayNet) and paper license terms when processing PDFs.