Athar Segmentation v4

Athar Segmentation v4 is the Kraken/BLLA line-segmentation model deployed with the Athar Arabic manuscript transcription system. It detects line regions and baselines before the separate exp9 recognition model transcribes each crop.

ملخص عربي

هذا نموذج القص واكتشاف أسطر المخطوطات العربية المستعمل في منظومة «أثر». النموذج لا يقرأ النص؛ بل يحدد مناطق الأسطر وخطوط الأساس، ثم يمرر القصاصات إلى نموذج التفريغ المستقل. دُرّب على 319 صفحة PAGE-XML من RASAM، وبلغ متوسط IoU=0.452 في سجل التدريب المحلي.

Evidence

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

The CER comparison measures the complete page pipeline, not segmentation alone. It shows the downstream impact of better crops, while IoU is the direct geometric segmentation metric. Dense or framed layouts can still require human PAGE-XML polygons.

Model file

Field Value
File model.mlmodel
Framework Kraken / BLLA
SHA-256 dc07cdeefe17598535ea3f5b187f7ddd2463baed909138c9ffd274c88b0e9a24
Intended input Full Arabic manuscript page image
Intended output Line polygons and baselines

Usage

Install a Kraken version compatible with the model and provide the model to Kraken's segmentation command. The Athar application loads it through:

SEG_MODEL_FILE=model.mlmodel

For scholarly work, inspect and correct the resulting line polygons before accepting a final transcription. Layout segmentation and recognition are separate sources of error.

Training data and license note

The recorded training source is RASAM, whose local dataset repository carries an Apache-2.0 license. This repository is initially kept private while the upstream warm-start model provenance and release terms are reviewed. The license: other metadata is deliberate; do not infer a permissive weight license until that review is complete.

Limitations

  • Framed, multi-column, marginal, and heavily overlapping layouts may fail.
  • The model is specialized for historical Arabic manuscript pages.
  • IoU does not measure reading order or transcription accuracy.
  • Human correction of PAGE-XML remains the safest path for complex pages.
Downloads last month
4
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support