docs(sources): add HTR tooling survey — Kraken/eScriptorium, MiDRASH, HTR4PGP, Eynollah, dhSegment

docs/sources/chatgpt_summary_htr_tools.md

@@ -0,0 +1,117 @@
+# HTR Tooling Survey — Layout Analysis and Line Segmentation for Hebrew Manuscripts
+
+*Ingested 2026-05-24. Source: ChatGPT survey on open-source tools for historical Hebrew manuscript
+layout analysis. Covers segmentation tools and pre-trained model projects relevant to the HASH
+corpus.*
+
+---
+
+## Context
+
+Standard layout analysis tools fail on historical Hebrew because they assume clean, LTR, printed
+text.  Historical Hebrew involves RTL writing, complex multi-column layouts (e.g. Talmud + Rashi
+commentary), marginalia, and heavily degraded parchment.  The tools below are the current
+best-practice open-source options used by digital humanities researchers and computer vision
+specialists for this material.
+
+---
+
+## 1. Kraken + eScriptorium (gold standard)
+
+**Kraken** is an open-source OCR/HTR engine optimised for historical documents and non-Latin
+scripts.  **eScriptorium** is its web-based graphical interface.
+
+- **RTL / BiDi native.**  Handles RTL and bidirectional text natively.
+- **Baseline tracing.**  Does not draw rectangular bounding boxes; traces the exact curvature of
+  handwritten baselines and generates polygon masks — essential for sloping or intersecting lines.
+- **Output formats.**  Exports ALTO-XML and PAGE-XML, the standard formats for HTR training datasets.
+- Kraken repo: <https://github.com/mittagessen/kraken>
+- eScriptorium repo: <https://github.com/scripta-studio/escriptorium>
+- Docker image available for self-hosted deployment.
+
+### Pre-trained Hebrew models (plug-and-play into eScriptorium)
+
+These projects have released open-source Kraken layout and text-recognition models trained
+specifically on hundreds of thousands of Cairo Genizah fragments.  They provide a high-accuracy
+first pass of baselines and bounding boxes without any training data collection:
+
+| Project | Training corpus | Model type | Location |
+|---------|----------------|------------|----------|
+| **MiDRASH Project** | Cairo Genizah (large) | Baseline segmentation + HTR | <https://github.com/MiDRASH-Project> (see releases) |
+| **Princeton Geniza Project — HTR4PGP** | Princeton Geniza Lab fragments | Baseline segmentation + HTR | <https://github.com/Princeton-CDH/geniza> (model registry) |
+
+**Relevance to this corpus:** The HASH corpus already includes Cairo Genizah fragments from
+Wikimedia Commons (Bodleian T-S items, Halper items) and has `openn__cairo_genizah` as a
+candidate source.  These models apply directly to that material.
+
+---
+
+## 2. Eynollah (Qurator-SPK)
+
+Command-line layout analysis tool using pixel-wise deep learning segmentation.
+
+- **Pixel classification:** segments a page into up to 10 classes: background, text region, text
+  line, header, image, separator, marginalia, table.
+- **Best use case:** highly complex physical layouts — e.g. a central Talmudic text surrounded by
+  dense wrapping Rashi commentary.  Eynollah is strong at detecting distinct structural zones
+  before drilling down to individual lines.
+- **Output:** fully structured PAGE-XML files with coordinates for every detected region and line.
+- Repo: <https://github.com/qurator-spk/eynollah>
+
+---
+
+## 3. LayoutParser
+
+Python library providing a unified API for deep learning document image analysis (Mask R-CNN,
+Faster R-CNN, etc.).
+
+- **Use case:** building an automated programmatic pipeline from scratch.  You annotate a few
+  dozen complex Hebrew pages, train a custom LayoutParser model, and run inference across
+  thousands of scans.
+- Repo / docs: <https://layout-parser.github.io/>
+
+---
+
+## 4. dhSegment (and Doc-UFCN)
+
+U-Net architecture for pixel-wise semantic segmentation.  Particularly robust against degraded,
+stained, or bleed-through manuscripts.
+
+- Instead of finding a "box," it classifies every single pixel as background, text line, or page
+  boundary.  Output is a mask that is easily converted to bounding boxes or polygons.
+- **Best use case:** the most damaged material in the corpus — heavily degraded Genizah fragments
+  and oldest Wikimedia Commons material.
+- dhSegment repo: <https://github.com/dhlab-epfl/dhSegment>
+- Doc-UFCN paper/repo: <https://github.com/soduco/paper-ufcn-icdar21>
+
+---
+
+## 5. HTR-United catalog
+
+A community catalog of published HTR training datasets and model registries.
+
+- Useful as a discovery tool: catalogs datasets that may include relevant Hebrew training material
+  not already tracked in this repo.
+- URL: <https://htr-united.github.io/catalog.html>
+
+---
+
+## Recommended workflow for annotation production (from survey)
+
+1. Set up **eScriptorium** via Docker.
+2. Load raw CC0 manuscript images from OPenn or Wikimedia.
+3. Apply the open-source **MiDRASH Geniza** baseline segmentation model to auto-trace lines.
+4. Use the eScriptorium UI to manually correct AI mistakes.
+5. Export corrected data as **PAGE-XML** to train a custom model.
+
+This workflow is directly compatible with the HASH corpus:  the OPenn candidate sources and the
+already-ingested Wikimedia Commons Genizah items are the right input material.
+
+---
+
+## Notes
+
+- The HASH entry schema already defines `alto_path` and `hocr_path` transcription fields; the
+  tools above are what would populate those paths in a future annotation phase.
+- This survey covers tooling only, not data sources.  All data-source leads remain in the other
+  `docs/sources/chatgpt_summary_*.md` and `gemini_summary_*.md` files.