--- language: - cs license: agpl-3.0 library_name: ultralytics tags: - cropilot - yolo - document-layout-analysis - document-cropping - digitization - vkol pipeline_tag: object-detection --- # VKOL Outer Crop v1 `vkol_outer_crop_v1.pt` is a Cropilot crop model trained for detecting outer page crop regions in VKOL digitization workflows (Olomouc Research Library, Czechia). The model is published here as a reusable model artifact. The surrounding `aicrop` project is a private VKOL implementation and deployment repository. It contains local Docker configuration, operational scripts, and institution-specific workflow glue, and is not currently published as a public dependency for this model. ## License This model is released under the GNU Affero General Public License v3.0 (`AGPL-3.0`). ## Model File - File: `vkol_outer_crop_v1.pt` - File size: 16,528,322 bytes - SHA256: `58ef487a285f3add1f937faeeec41775aaddca75211083fc49ec8decf54d17a9` The filename follows the Hugging Face repository name. In VKOL's private deployment the same model may be installed under an operational Cropilot model name such as `outer_crop.pt`. ## Intended Use The model is intended for semi-automated crop detection in a Cropilot/Ořezy workflow: 1. Working JPEG derivatives are created from original TIFF scans. 2. Cropilot runs AI detection using this crop model. 3. A human operator reviews and corrects the detected coordinates in the Cropilot editor. 4. Final crops are applied to the original TIFF scans by the local production tooling. The model predicts crop regions only. It does not create final archival TIFF files and does not perform color management, metadata handling, compression, or output delivery. ## Recommended Integration Typical VKOL/Cropilot settings: - Crop model: `outer_crop` or another local alias pointing to this file - Rotation model: `text` - Input for detection: downscaled working JPEG derivatives - Final production input: original TIFF scans ## Training Data The model was trained or fine-tuned from VKOL Cropilot annotations and/or local YOLO datasets prepared from digitization material. The training data is not included in this repository. Because the training data reflects local production material and annotation conventions, the model should be treated as domain-specific rather than a general-purpose document layout model. ## Limitations - Detection quality depends on scan type, derivative resolution, contrast, skew, page borders, and local annotation conventions. - The model may not generalize well to unrelated collections without validation or further fine-tuning. - Human review in Cropilot is recommended before producing final archival output. - Final image conversion steps should be handled by the downstream production pipeline. ## Example Download ```python from huggingface_hub import hf_hub_download path = hf_hub_download( repo_id="bezverec/vkol_outer_crop_v1", filename="vkol_outer_crop_v1.pt", ) print(path) ``` ## Upstream Context The model is designed for the Cropilot/Ořezy ecosystem developed around document crop detection workflows. Relevant upstream public components include: - - - VKOL's `aicrop` repository is a private implementation/deployment layer around these components. ## Versioning This repository stores the v1 outer crop model used in the VKOL workflow. Newer fine-tuned models should be uploaded as separate commits, files, or repositories so deployments can remain reproducible.
