File size: 3,957 Bytes
6d1698d | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 | ---
language:
- bo
license: agpl-3.0
library_name: ultralytics
pipeline_tag: image-segmentation
datasets:
- openpecha/OCR-Tibetan_line_segmentation_coordinate_annotation
tags:
- tibetan
- document-ai
- line-segmentation
- instance-segmentation
- yolo11
- ultralytics
---
# PechaBridgeLineSegmentation
YOLO11n instance-segmentation model for detecting individual Tibetan text
lines on traditional pecha page scans. The model is the line-layout stage used
by [PechaBridge](https://github.com/CodexAITeam/PechaBridge) before OCR.
> **Important:** The supported PechaBridge inference pipeline applies the
> `gray` preprocessing mode (minimum RGB channel, no binarization) before YOLO
> inference. Raw Ultralytics inference can produce different results.
## Recommended usage — PechaBridge CLI
```bash
git clone https://github.com/CodexAITeam/PechaBridge.git
cd PechaBridge
pip install -r requirements.txt
# Downloads this model and the PechaBridge OCR model.
python cli.py download-models
python cli.py batch-ocr \
--input-dir /path/to/pecha/pages \
--ocr-model models/ocr/PechaBridgeOCR \
--line-model models/line_segmentation/PechaBridgeLineSegmentation.pt \
--layout-engine yolo_line \
--line-preprocess gray \
--ocr-engine donut
```
## Python usage
```python
from pathlib import Path
from huggingface_hub import snapshot_download
from PIL import Image
from ultralytics import YOLO
from pechabridge.ocr.line_segmentation import (
apply_line_segmentation_preprocess,
)
model_dir = Path(snapshot_download(
"TibetanCodexAITeam/PechaBridgeLineSegmentation"
))
weights = next(model_dir.glob("*.pt"))
model = YOLO(str(weights))
image = Image.open("page.jpg").convert("RGB")
prepared = apply_line_segmentation_preprocess(image, pipeline="gray")
results = model.predict(
source=prepared,
imgsz=1280,
conf=0.25,
verbose=False,
)
for result in results:
print(result.boxes.xyxy) # line bounding boxes
print(result.masks.xy) # line polygons
```
## Model details
- **Architecture:** Ultralytics YOLO11n segmentation
- **Task:** single-class instance segmentation
- **Class:** `line`
- **Weights:** `yolo_line_seg.pt` (about 6.0 MB)
- **Recommended input size:** `1280`
- **Recommended confidence threshold:** `0.25`
- **Recommended preprocessing:** PechaBridge `gray`
- **Ultralytics version used for the exported checkpoint:** `8.4.14`
## Training data
The training corpus was derived from
[`openpecha/OCR-Tibetan_line_segmentation_coordinate_annotation`](https://huggingface.co/datasets/openpecha/OCR-Tibetan_line_segmentation_coordinate_annotation),
converted to YOLO polygons, padded, filtered, and split locally:
| Split | Images | Label files |
|---|---:|---:|
| Train | 3,487 | 3,487 |
| Validation | 519 | 519 |
| Test | 503 | 503 |
## Validation results
Metrics stored in the released checkpoint:
| Output | Precision | Recall | mAP50 | mAP50–95 |
|---|---:|---:|---:|---:|
| Bounding boxes | 0.97736 | 0.96768 | 0.99138 | 0.61298 |
| Segmentation masks | 0.90993 | 0.90055 | 0.89697 | 0.41885 |
These numbers describe the local validation split and should not be treated as
an independent cross-collection benchmark.
## Intended use and limitations
- Intended for locating horizontal Tibetan text lines in traditional pecha scans.
- Not an OCR model; detected line crops must be passed to a text recognizer.
- Performance may degrade on modern book layouts, handwriting, vertical text,
heavy page curvature, severe blur, unusual ornaments, or unseen collections.
- Closely spaced or touching lines may be merged; damaged lines may be fragmented.
- Predictions should be reviewed before scholarly or archival publication.
## License
AGPL-3.0. The checkpoint was trained with Ultralytics YOLO, whose open-source
software and trained model weights are distributed under AGPL-3.0 by default.
Commercial or closed-source use may require an Ultralytics Enterprise License.
|