Domain-Adapted YOLO Line Segmenter — Historical Documents
🔗 Live visual demo: try the segmenter on real historical pages (original vs detected lines) at https://huggingface.co/spaces/abhishekjha1008/historical-line-segmenter-demo
A YOLOv8m line detector, domain-adapted to historical document layouts, for the segmentation stage of an HTR pipeline. Segmentation is the dominant hidden error source in historical HTR: adapting the segmenter to the target layouts recovers most of the gap before a recogniser runs.
| Metric | Value |
|---|---|
| mAP50 | ~0.79 |
Usage
from ultralytics import YOLO
model = YOLO("yolov8m_lineseg_v2_best.pt")
results = model("page.jpg") # each box = one text line; crop and pass to a recogniser
A box-based counterpart to the baseline-based kraken segmenter. Related: synthetic dataset · live demo.
License
Built on Ultralytics YOLOv8 (AGPL-3.0) — respect the upstream license.
Author: Abhishek Jha · github.com/AbhiPandit1 · hf.co/abhishekjha1008