Toward a Born-Digital Edition: Reflowing Scanned Technical Books¶

Status. Experiment report, 2026-11-06. Reproducible from experiments/reflow/ (see its README). The code is not wired into the pipeline yet; see §8 for the proposed integration.

Question. How close can the digitizer get to a document that looks born digital, rather than a cleaned-up scan? If possible, the document should be rebuilt from scratch, for example as a reflowable EPUB.

Test book. Fyzika po kapitolách 1: Vektory (Ivan Červeň, STU 2007, ISBN 978-80-227-2663-4), fei_CVI_OPACID_FEI_9788022726634:

• 58 pages, of which 53 content pages were processed;
• a hard case: dense vector calculus, determinants, line-art figures, glossary, library stamps.

Input caveat. The raw-scan cache was gone, so every experiment ran on the cached legacy-engine output. That is RGB JPEG at about 300 dpi, with bleed-through and a facing-page sliver on some pages. §7 checks the analyzer on new-engine (ScanTailor, 600 dpi) pages.

All quality judgements below are visual checks by the author, not measurements against a ground-truth transcription.

1. Summary¶

EPUB validation.

• Variant B was checked with epubcheck 5.4.0. It reported 6 errors, all OPF-073: potrace writes an SVG DOCTYPE. The fix is in the code.
• Variant C has not been re-validated, because the re-run was blocked in the session. Do not assume C passes until epubcheck (and ideally Ace) have run.

Recommendation. Ship C as an additional, derived edition (<slug>.epub). Keep the MRC PDF as the faithful master. C is the only variant that reads like a born-digital book: it reflows, its math is selectable and accessible (MathML), it is about 15× smaller, and it can be re-typeset to PDF.

C still has visible residual errors (§6), so it is a reading edition, not a replacement for the facsimile.

2. Literature and landscape (condensed)¶

Full-page VLM parsers. As of late 2026, the leaders on OmniDocBench v1.6 are PaddleOCR-VL-1.6 (96.3), MinerU2.5-Pro (95.8) and GLM-OCR (95.2). These numbers are self-reported or come from aggregator leaderboards; they were not reproduced here.

• PaddleOCR-VL is the only one that is all of: Apache-2.0 code and weights, 0.9B parameters (runs on a laptop), and with Slovak named in its language list.
• Marker/Chandra and Surya ship OpenRAIL-M weights with commercial restrictions.
• Granite-Docling and GLM-OCR do not cover Slovak.
• olmOCR-2 needs an NVIDIA GPU.

Hallucination literature. Small VLMs "rewrite" toward their language priors (ACL 2026 Linguistic Priors Hallucination; Do VLMs Read or Rewrite? 2026). These errors are invisible to CER (When Low CER is Not Enough, 2026). LLM post-correction gives no gain for lower-resource languages (No Free Lunches, 2025). Correction works when it is anchored to a reference (AAAI 2025).

ClearScan-style vector glyphs. No mature open-source implementation exists. ScanWeave is early-stage, and smoothscan has weak results.

Practice.

• Distributed Proofreaders and Wikisource rely on human page proofing.
• Standard Ebooks starts from existing transcriptions.
• EPUB accessibility (EPUB A11y 1.1, legally relevant since the EAA in June 2025) favours reflowable EPUB over fixed-layout page images.

YoGaDoc (an STU project) contributes one transferable idea: DocLayout-YOLO (DocLayNet classes) as a trained layout detector. Its EasyOCR is weaker than Tesseract slk for Slovak. DocLayout-YOLO is AGPL-3.0. It was not run (§9).

3. Architecture of variant C¶

page image + Tesseract slk hOCR
  └─ analyze: blocks (heading | para | list | table | formula | figure | caption)
       ├─ prose        -> Tesseract text (diacritic authority)
       ├─ 2-D math     -> PaddleOCR-VL "Formula Recognition:" -> LaTeX
       ├─ prose+math   -> PaddleOCR-VL "OCR:" -> anchored to Tesseract words
       ├─ tables       -> rebuilt from Tesseract word geometry (no VLM)
       └─ figures      -> crop, paper->white, line art traced to SVG (potrace)
  └─ guards -> Markdown -> pandoc: EPUB 3 (MathML) + xelatex PDF

Routing is per line, on Tesseract signals:

• Low word confidence with few dictionary-like words marks math. Prose scores 93–96 and display math 38–43.
• Inline-math characters, or Tesseract's misread-Greek signature (@ © ® for ω), mark a mixed line.

Anchoring. VLM prose tokens snap to Tesseract's high-confidence words. The snapping is:

• order-aware: anchors are consumed left to right, within a look-ahead window;
• diacritic- and script-insensitive: funkcίου≡funkciou, polá≡poľa;
• arbitrated by the Slovak hunspell dictionary: if only the VLM word is a real word, it wins, as in jednotkový.

4. Negative result 1: a SOTA VLM on Slovak prose¶

Here is PaddleOCR-VL-1.6 OCR: on a clean prose block that Tesseract slk reads perfectly. The excerpt is verbatim.

…vektorovou funkcίου. … súvislost … kładných … zdroje polá … V magnetických políach, kde indučněnéčiary sú … v úbobovnom bode … súradnicovej sústav

The output contains:

• Greek letters inside a Slovak word;
• Polish ł and Czech ě;
• dropped ť/ľ;
• two words merged into one;
• a truncated ending.

This reproduces the 2026-06 PERO-OCR finding with a different, newer model: benchmark leadership does not transfer to Slovak prose. The VLM is therefore never trusted for prose.

5. What works: math recovery¶

Formulas. On visual inspection (experiments/reflow/formula_sheet_40.png), display formulas come out as correct LaTeX, including 3×3 determinants, multi-line systems and long expansions. Typical results:

• f=\sqrt{f_x^2+f_y^2+f_z^2}
• the full rot F determinant expansion

Anchoring on inline/mixed lines. 136 segments, 480 anchors:

The "foreign letters" column is counted before the per-line guard. Any line that still contains a non-Slovak letter after anchoring falls back to Tesseract. A scan of the final Markdown finds 0 such letters in prose.

On these hard lines the hybrid recovers more Slovak-only diacritics than either engine alone.

Guards.

6. Residual errors (visual check)¶

• Rare, silent formula misreads. Examples: div read as \mathrm{l i v}, and ∂ read as λ with an extra determinant row. They are syntactically valid, so no guard catches them. This is the main open risk for a physics text.
• Mitigation (future work): a second formula recognizer as a cross-check, or render-and-compare against the crop.
• Greek misreads in high-confidence prose lines are not routed. Examples: Ar for Δr, r» for r₂.
• Display equations sometimes typeset inline (dr/dt = lim…). These are the mixed-line path.
• Figure labels outside the detected box leak into the text. Captions get a doubled prefix (Obrázok 7: Obr. 1.3.2.1).
• The re-typeset PDF is not hyphenated. TeX Live basic lacks the Slovak patterns.
• Variant B is no smaller than A0. Its formula crops are unoptimized PNGs.

7. Transfer to new-engine input (ScanTailor mixed, 600 dpi)¶

The analyzer ran on 600 dpi grey renders of the new-engine fad distribution PDF (JPEG background plus JBIG2 mask). That is a proxy for ScanTailor output, not the actual ScanTailor TIFFs, which were not cached.

• Tables (nie / áno value columns) and lettered lists classify correctly.
• The running head is no longer removed. ScanTailor margins push it below the 10 % band. This needs book-level detection of repeated text.
• Worksheet/form pages become noise. This needs a whole-page image fallback.
• Colour stamp removal cannot work on bitonal mixed output. It would have to run before ScanTailor. Its production placement is still to be decided.
• fad has no math, so the math path is validated only on legacy input.

8. Vector glyphs (A1): cosmetic only¶

Tracing the 600 dpi JBIG2 mask with potrace gives 282–361 KiB of vector outlines per page. At reading zoom the edges are smoother and specks vanish. But the glyph bumps from the 300→600 dpi upsample are traced faithfully, and each a stays slightly different (experiments/reflow/raster_vs_vector.png).

A born-digital look needs either glyph clustering (identical glyph instances, a Type 3 font) or full re-typesetting. Glyph clustering is risky for Slovak diacritic pairs (l/ľ, c/č).

9. Throughput and operations¶

Prose-only books need almost no VLM calls. Corpus-scale math needs the official GGUF/vLLM serving on a GPU; HF generate on MPS runs at about 8 tok/s.

Operational gotchas found.

• Pathological crops (3 × 140 px) explode under the processor's min-pixel upscaling and OOM the GPU. Crops are now white-padded to aspect ≤ 10 and capped.
• MPS memory accumulates across generate calls, and torch.mps.empty_cache() does not release it. The driver's "other allocations" grow to about 40 GB until jobs OOM. In the first pipeline end-to-end run, 30 of 67 jobs failed this way. The runner therefore recycles the worker process every 25 jobs and retries an OOM'd job once in a fresh process (reflow/vlm.py: JOBS_PER_PROCESS, MAX_ATTEMPTS).
• The VLM result cache is content-addressed (page + bbox + task), so re-analysis never mis-attaches stale results.

10. Integration (2026-11-07)¶

The edition is now built in the pipeline by the opt-in step ReflowEdition (pipeline/steps/reflow.py). The code lives in the evilflowers_books_digitalizer/reflow/ package:

Pipeline placement. The step runs after EnsurePdfA, while the cleaned pages are still in the work dir. Its input is the pipeline's multi-page hOCR at 600 dpi. Its output is <slug>.epub, plus a report in metadata["reflow"]. It is fail-safe: an error is recorded, but the book is never failed.

In the Restoration Studio. The studio's renderer (studio/adapters/render) builds <id>.epub with the same reflow.edition.build_edition, from the access leaves and the book's hOCR. It does not yet use the studio's region decisions or VLM answers: formulas in the studio EPUB are always images, even when an operator confirmed their LaTeX. Feeding the confirmed regions into the edition is the next step.

Configuration.

[reflow]
enabled = false          # opt-in
vlm = "none"             # or "paddleocr-vl"
# vlm_python / vlm_model / vlm_device / hunspell_dict — see configs/pipeline.toml

Tool paths can also be overridden by $EVILFLOWERS_VLM_PYTHON and $EVILFLOWERS_VLM_MODEL. preflight checks for pandoc and, as optional tools, potrace and hunspell when reflow is enabled.

Both §7 prerequisites are implemented.

• Running heads are detected per book: a short line in the top band that repeats across the book.
• Form or worksheet pages become a single whole-page image. The trigger is a low word-like ratio together with mostly fragment paragraphs.

Math validation changed. It is now one batched pandoc → MathML run per book, instead of an xelatex compile per snippet, because EPUB is the shipped format.

Still deferred:

• the re-typeset LaTeX PDF;
• colour stamp removal before ScanTailor (the analyzer removes coloured stamps only when colour survives into the page);
• a trained layout source such as DocLayout-YOLO.