factlogic/athar-segmentation-v4

Athar Segmentation v4

Athar Segmentation v4 is the Kraken/BLLA line-segmentation model used in the Athar Arabic manuscript transcription workflow.

It detects manuscript line regions and baselines before the separate Phoenix HTR model performs text recognition.

Paper

The Athar workflow and Phoenix recognizer are described in:

Beyond Recognition: Compact Multi-Domain Arabic Manuscript HTR with Candidate-Selection Analysis and Evidence-Preserving Review

arXiv:2608.19385

https://arxiv.org/abs/2608.19385

ملخص عربي

Athar Segmentation v4 هو نموذج اكتشاف وتقسيم أسطر المخطوطات العربية المستخدم ضمن منظومة أثر Athar.

النموذج لا يتعرف على النص نفسه، بل يحدد:

• مناطق الأسطر
• خطوط الأساس Baselines
• القصاصات التي تُرسل لاحقًا إلى نموذج Phoenix

دُرّب النموذج على 319 صفحة PAGE-XML من RASAM، وبلغ متوسط:

IoU = 0.452

في سجل التدريب المحلي.

Results

Measurement	Result
Training pages	319 RASAM pages
Recorded mean IoU	0.452
Full-pipeline CER with older segmenter	10.06%
Full-pipeline CER with v4	8.43%

The CER values come from a 30-page end-to-end page-pipeline comparison.

They measure the combined effect of segmentation and downstream recognition, not segmentation accuracy alone.

The IoU value is the direct geometric segmentation metric.

Model details

Field	Value
Framework	Kraken / BLLA
Model file	model.mlmodel
Input	Full Arabic manuscript page image
Output	Line polygons and baselines
SHA-256	dc07cdeefe17598535ea3f5b187f7ddd2463baed909138c9ffd274c88b0e9a24

Usage

The Athar application loads the model using:

SEG_MODEL_FILE=model.mlmodel

The segmentation output can then be passed to Phoenix for recognition.

For difficult manuscript layouts, PAGE-XML regions should be inspected and corrected manually before producing a final scholarly transcription.

Intended use

This model is intended for:

- Arabic manuscript line segmentation
- baseline detection
- PAGE-XML assisted transcription workflows
- preprocessing for Arabic HTR systems
- human-in-the-loop manuscript digitization

Limitations

- Framed and multi-column pages may be difficult.
- Marginalia and overlapping lines may require manual correction.
- The model is specialized for historical Arabic manuscript layouts.
- IoU does not measure reading order or transcription correctness.
- Segmentation errors can propagate into the recognition stage.

Training data and license

The recorded training source is RASAM.

The retained local dataset repository records RASAM under Apache-2.0.

However, the complete provenance and release terms of the upstream warm-start segmentation model have not yet been fully verified. For this reason, this repository uses:

"license: other"

This should not be interpreted as a permissive model-weight license until the provenance review is complete.

Citation

If you use Athar or Phoenix in research, please cite:

@article{ali2026beyond,
  title   = {Beyond Recognition: Compact Multi-Domain Arabic Manuscript HTR with Candidate-Selection Analysis and Evidence-Preserving Review},
  author  = {Ali, Abdullah Ahmed and Abdulhadi, Mohammed Thamer and Safaa, Ali Haider and Wadi, Dhulfiqar Mahdi},
  journal = {arXiv preprint arXiv:2608.19385},
  year    = {2026}
}

Scholarly note

Segmentation and recognition are separate sources of error.

For complex manuscript pages, the safest workflow is:

page image -> segmentation -> human polygon review -> Phoenix recognition -> human transcription review