Pagescan Weights

--- license: apache-2.0 tags: - document-scanner - document-detection - perspective-correction - object-detection - segmentation - yolo - sam library_name: pagescan pipeline_tag: object-detection --- # pagescan weights Pre-trained model weights used by [pagescan](https://github.com/7RPlus-GmbH/pagescan), a privacy-first document scanner that turns phone photos into clean, deskewed, print-ready PDFs. These weights power the default detection cascade: 1. **YOLO11** detects a coarse axis-aligned bounding box around the document. 2. **HQ-SAM ViT-B**, prompted by that bbox, returns a precise binary mask. 3. The mask is fitted to a 4-corner quadrilateral for perspective correction. 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 orient...

Source context: 0 downloads · 1 likes · Pipeline object-detection · Library pagescan · Repo 7rplus/pagescan-weights