strikethrough-ocr

Most OCR pipelines silently drop text decorations like color, underline, strike-through.

StrikethroughOCR tries to preserves them.

Installation

uv tool install strikethrough-ocr    # or: pip install strikethrough-ocr

Tesseract itself must be on PATH:

brew install tesseract               # macOS
apt install tesseract-ocr            # Debian/Ubuntu

Command line

strikethrough-ocr [OPTIONS] INPUT.png

Without options it writes the HTML overlay next to the image:

strikethrough-ocr page.png                 # -> page.scan-preview.html (overlay)
strikethrough-ocr -to html page.png        # -> page.scan.html (clean document)
strikethrough-ocr -to md page.png          # -> page.scan.md
strikethrough-ocr page.png -o out/         # choose the output directory
strikethrough-ocr page.png --stdout        # print instead of writing
strikethrough-ocr page.png --lang deu+eng  # Tesseract language(s)
strikethrough-ocr page.png --json          # also write page.scan.json
strikethrough-ocr page.png --debug         # also write page.debug.png
strikethrough-ocr page.png -v              # log second-pass OCR decisions

The HTML overlay references the scan by relative path for its slider, so keep the two files together or the slider shows nothing.

Library

from strikethrough_ocr import process, render_html, html_to_markdown, \
    html_to_document_html

model = process("page.png", lang="eng")           # run the OCR pipeline
overlay = render_html(model, "page", "page.png")  # styled HTML overlay
md = html_to_markdown(overlay)                    # Markdown in reading order
doc = html_to_document_html(overlay)              # clean document HTML

process() returns a PageModel: words (text, box, baseline, font size, colour, bold, underline, strike, decoration colour, confidence), table rules, decoration strokes and graphics. model_to_json() turns it into plain records; draw_debug() paints the detections onto the scan image.

Development

uv sync --group dev          # install with the test dependencies
uv run pytest                # unit tests and golden files
UPDATE_GOLDENS=1 uv run pytest   # rewrite tests/golden and the overlay fixture

The golden tests pin all three renderings of one synthetic fixture page (tests/conftest.py builds it): the overlay it renders must equal tests/fixtures/page.overlay.html, and the Markdown and document HTML produced from that fixture must equal the files under tests/golden/. After a deliberate change to a renderer, regenerate them and review the diff. One end-to-end test runs the real pipeline and is skipped when no tesseract binary is on PATH.

How it works

1. Ink mask. The page is thresholded on each pixel's distance from white, 255 − min(B, G, R), so black and coloured ink are both caught. The glyph size measured on this mask sets the working scale: a scan whose print is smaller than the 300 dpi reference is enlarged (up to 4x) before anything else, because Tesseract and every stroke measure need the pixels. Highlighter bands pass the ink threshold too; they are recognised as bright, broad fills — every print ink is dark, and no stroke of print is as tall as a marker band — and whitened away, so OCR sees paper while the letters on top of them stay.
2. Straight strokes. Morphological opening finds long horizontal and vertical strokes. A stroke with clear paper on both sides is a table rule; every other horizontal stroke is a decoration candidate (underline or strike-through). The paper test follows the stroke's ink column by column, so a slightly slanted strike-through does not pass for a rule. A stroke crossed by vertical rules belongs to a table lattice: it is split at the crossings, and only a piece whose glyphs pass through it (a strike drawn flush onto a row boundary fuses with it into one line) stays a decoration. Length is deliberately no criterion — a struck-through paragraph line is as long as a rule.
3. OCR. Strikes are removed in two ways (Telea inpainting and column-wise erasure, which fail on different glyphs) and Tesseract reads the cleaned page as hOCR, which carries a baseline and font size per row. Weak lines are cut out and read again in isolation.
4. Per-word style. Colour is the median of the darkest ink in the word box (darkest by luminance, so the fringe where a glyph blends into a marker or the paper does not vote); a decoration's colour comes from the stroke's own pixels, or the glyphs a red line crosses would outvote the line. Bold compares the measured stroke width (2 × area / perimeter of the word's core ink) against a regular-weight model fitted to the page. Decoration strokes are assigned to the words they span: through the x-height it is a strike, at the baseline an underline. Struck lines are re-read on both cleaned variants and the most confident reading wins.
5. Graphics. Oversized coloured lettering (logos) and ink no confident word claims are cropped from the scan and embedded as PNG data URIs.

In the Markdown output, connected rules become tables (merged cells hold their content in the top-left grid cell), large text becomes headings, bold is **text**, strike-through ~~text~~, underline <u>text</u> and colour <span style="color:#rrggbb">…</span>; colour and underline have no Markdown syntax, so they are inline HTML. The document HTML output shares the same layout analysis and writes it the way pandoc would: <p>, <h1..h3> with an id, <table> with a <colgroup>, <strong>/<del>/<u>/<span> inline, and consecutive paragraphs labelled (vi), (b) or 3. folded into <ol start type> lists — unless the label itself is struck through, because deleted text keeps its literal label.

Limits and assumptions

• Geometry constants are set for a 300 dpi scan and scaled to the page's measured glyph size; a scan with smaller print is enlarged internally (up to 4x) first. Below roughly 8 px of x-height there is not enough signal left even so, and the bold measure goes first.
• The page is assumed to be deskewed; a rotated page breaks the rule detection.
• Bold is relative: a page set entirely in bold has no regular reference and renders as regular. Italic is not detected.
• Underline and strike-through are found as straight strokes at least 40 px long (at 300 dpi); a wavy or hand-drawn line is missed, and a heavy or doubled strike hides too much of the glyphs for the text to be recovered. A marker stroke as tall as the text it covers is taken for a highlight, not a strike.
• The HTML overlay is a reproduction of the page, not a document: words are absolutely positioned and the reading order is Tesseract's. The document HTML and Markdown outputs are the document view.