license: agpl-3.0
language:
- cs
library_name: ultralytics
pipeline_tag: object-detection
tags:
- cropilot
- yolo
- document-layout-analysis
- document-cropping
- digitization
Default Crop v1
default_crop_v1.pt is a Cropilot model for detecting page crop regions in scanned printed documents. It is the general-purpose default crop model, trained on a mixed corpus of books, magazines, and newspapers.
It is published here as a reusable model artifact for use in the Cropilot document-cropping ecosystem.
License
Released under the GNU Affero General Public License v3.0 (AGPL-3.0).
Model File
- File:
default_crop_v1.pt - Base model: YOLO11s
- Task: object-detection (page crop region)
Intended Use
This model is intended for semi-automated processing in a Cropilot workflow:
- Compressed JPEGs are created from original scans (e.g. TIFF).
- Cropilot runs AI detection using this model.
- A human operator reviews and corrects the result in the Cropilot editor.
- Final crops are applied to the original scans by the production tooling.
The model handles both single-page and double-page (spread) scans, making it a good general default when the document type is mixed or unknown.
Recommended Integration
Typical Cropilot settings for this model:
- Crop model:
default_crop - Rotation model:
text
Training Data
The model was trained on a mixed corpus of digitized printed documents covering both single-page and double-page layouts. Approximate composition:
- ~80% books
- ~15% magazines / periodicals
- ~5% newspapers
The material spans roughly 1950–2010.
Limitations
- Performance is strongest on recent books; older material, heavily illustrated layouts, or document types outside the training mix may need validation or further fine-tuning.
- Book covers are detected poorly.
Example Download
from huggingface_hub import hf_hub_download
path = hf_hub_download(
repo_id="cropilot-community/default_crop_v1",
filename="default_crop_v1.pt",
)
print(path)
Versioning
This repository stores v1 of the default crop model. Newer fine-tuned models should be published as separate v<n> repositories so deployments remain reproducible.