This project develops a layout parsing pipeline to extract key components (e.g., abstract, context, table, reference) from academic PDFs using a Detectron2-based model trained on annotations from Label Studio.
To identify and segment document elements like titles, authors, abstracts, tables, figures, and references using object detection techniques, improving downstream analysis and semantic classification with LLMs.
- Fast R-CNN architecture (Detectron2) for layout detection
- Layout categories: Abstract, Author, Context, Header, Image, Reference, Sub-title, Table, Title
- Integration-ready with LLMs for content-based filtering or labeling
- Configuration through
config.yaml
config.yaml- Detectron2 configuration for the layout modelresult.json- Output annotations from model inferenceparsing.ipynb- Sample notebook to run detection and visualize results
Install via pip:
!pip install pycocotools !pip install layoutparser !pip install "layoutparser[effdet]" !pip install layoutparser torchvision !python -m pip install 'git+https://github.com/facebookresearch/detectron2.git' !pip install layoutparser torchvision && pip install "git+https://github.com/facebookresearch/detectron2.git@v0.5#egg=detectron2" !pip install "layoutparser[paddledetection]" !pip install "layoutparser[ocr]"
Install via Conda:
conda install detectron2 pytorch opencv omegaconf hydra-core -c conda-forge
π How to Run
# Inside parsing.ipynb from layoutparser.models import Detectron2LayoutModel model = Detectron2LayoutModel( config_path='config.yaml', extra_config=["MODEL.ROI_HEADS.SCORE_THRESH_TEST", 0.8], label_map={0: "Abstract", 1: "Author", ...} )
π Annotation Categories β’ Abstract β’ Author β’ Context β’ Header β’ Image β’ Reference β’ Sub-title β’ Table β’ Title