A legacy fallback chain is also bundled, used when the cascade is unavailable (e.g. no torch) or fails on a given image: DocAligner's FastViT-SA24 and LCNet100 corner-heatmap regressors, plus a DeepLabV3-MobileNetV3 document-segmentation net for the conservative-crop path. These are re-hosted third-party weights — see attribution below.

Files

File	Size	Description
`yolo_doc_v1.onnx`	11 MB	Exported YOLO11 detector — used at inference time by `pagescan.detector` (no torch required).
`yolo_doc_v1.pt`	5.3 MB	Original PyTorch checkpoint — kept for re-export / fine-tuning.
`sam_hq_vit_b.pth`	362 MB	HQ-SAM ViT-B checkpoint — re-hosted copy (see attribution below).
`fastvit_sa24_h_e_bifpn_256_fp32.onnx`	83 MB	DocAligner FastViT-SA24 corner-heatmap regressor — legacy primary (re-hosted; see attribution).
`lcnet100_h_e_bifpn_256_fp32.onnx`	4.8 MB	DocAligner LCNet100 corner-heatmap regressor — legacy fallback backbone (re-hosted; see attribution).
`deeplabv3_mbv3_docseg.onnx`	44 MB	DeepLabV3-MobileNetV3 document segmentation — conservative-crop fallback (re-hosted; see attribution).

Usage

These weights are downloaded automatically by pagescan on first use:

bash

pip install pagescan
python -c "from pagescan import scan; scan('photo.jpg', 'out.pdf')"

To pin a specific revision or use a local cache:

python

from huggingface_hub import hf_hub_download

onnx_path = hf_hub_download(
    repo_id="7rplus/pagescan-weights",
    filename="yolo_doc_v1.onnx",
)

Training

yolo_doc_v1 was trained from yolo11n-obb.pt on a 1000-photo private corpus of phone-captured documents (Dec 2025) with oriented bounding box (OBB) labels. The training script lives in training/yolo/. A v2 trained on an extended, distribution-balanced corpus is in progress.

Attribution — HQ-SAM

sam_hq_vit_b.pth is not original work from this repository. It is a re-hosted copy of the ViT-B HQ-SAM checkpoint from the HQ-SAM authors, included here so pagescan installs ship with a single, stable weight source.

HQ-SAM — Segment Anything in High Quality Lei Ke, Mingqiao Ye, Martin Danelljan, Yifan Liu, Yu-Wing Tai, Chi-Keung Tang, Fisher Yu. NeurIPS 2023. Paper: https://arxiv.org/abs/2306.01567 Code & original weights: https://github.com/SysCV/sam-hq Original Hugging Face mirror: https://huggingface.co/lkeab/hq-sam

HQ-SAM is released under the Apache 2.0 license. We re-host the ViT-B checkpoint unchanged; all credit for the model itself belongs to the original authors. If you use pagescan in research that depends on the cascade's segmentation quality, please cite the HQ-SAM paper:

bibtex

@inproceedings{ke2023segment,
  title     = {Segment Anything in High Quality},
  author    = {Ke, Lei and Ye, Mingqiao and Danelljan, Martin and Liu, Yifan and Tai, Yu-Wing and Tang, Chi-Keung and Yu, Fisher},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2023}
}

Attribution — DocAligner (legacy weights)

fastvit_sa24_h_e_bifpn_256_fp32.onnx and lcnet100_h_e_bifpn_256_fp32.onnx are not original work from this repository. They are re-hosted copies of the corner-detection ONNX weights from the DocAligner project by DocsaidLab, included so pagescan's legacy fallback path has a single, stable weight source (the DocAligner upstream is an undocumented auto-downloader).

DocAligner — DocsaidLab Code & docs: https://github.com/DocsaidLab/DocAligner · https://docsaid.org/en/docs/docaligner/

DocAligner is released under the Apache 2.0 license. We re-host these checkpoints unchanged; all credit for the models belongs to DocsaidLab.

deeplabv3_mbv3_docseg.onnx is a re-hosted DeepLabV3-MobileNetV3 document-segmentation model used only by the conservative-crop fallback. Its upstream provenance is not definitively traced; if you redistribute it, confirm its original source and license.

License

yolo_doc_v1.{pt,onnx} — released under Apache 2.0 by 7R+ GmbH.
sam_hq_vit_b.pth — Apache 2.0, HQ-SAM authors (see above).
fastvit_sa24_*.onnx, lcnet100_*.onnx — Apache 2.0, DocAligner / DocsaidLab (see above).
deeplabv3_mbv3_docseg.onnx — re-hosted; provenance/license to be confirmed.

The pagescan package itself is MIT-licensed; see the main repository.