Image Colour Contrast Checker

Live tool: image-colour-contrast-checker.timdixon.net

A privacy-respecting, fully client-side web app that audits images for text accessibility. Drop in one or more screenshots, get back a per-image report: for every foreground/background colour pair it runs six checks — WCAG AA, WCAG AAA, APCA, colour-vision-deficiency contrast, vestibular saturation and a cognitive verdict — alongside whole-image colour-blindness simulations, cropped failing regions, and one-click PDF or Markdown export.

Nothing is uploaded. OCR, contrast analysis, and report generation all run in your browser. See the Privacy Statement for full details.

For a full technical description of every module, the data pipeline, and how to extend the tool, see ARCHITECTURE.md.

---

Features

• WCAG 2.2 AA & AAA contrast analysis with the per-strip "worst-case" sampling used by accessibility auditors (catches gradient backgrounds where contrast degrades mid-word).
• APCA perceptual contrast — the Advanced Perceptual Contrast Algorithm, vendored from apca-w3 into src/core/apca.js (no runtime dependency). A beta perceptual signal beyond the WCAG ratio (not a WCAG requirement).
• Colour-vision-deficiency contrast — each pair's contrast recomputed for deuteranopia, protanopia and tritanopia using the Machado (2009) matrices, flagging pairs that pass for normal vision but fail for a CVD viewer.
• Whole-image colour-blindness simulation — every image rendered as four common deficiencies see it.
• Vestibular & cognitive checks — flags highly saturated pairings that can shimmer, and rolls every check into one plain-language verdict.
• Expandable per-combination report — one row per colour pair, collapsed until you open it. A static "What the checks mean" section explains every check, and each check's ⓘ link jumps straight to its entry.
• PaddleOCR PP-OCRv4 text detection + recognition via @gutenye/ocr-browser + onnxruntime-web. WebGPU on Chrome/Edge desktop; WASM fallback everywhere else.
• k-means colour split per OCR region — text = minority cluster, background = majority — so it works on real-world UI screenshots, not just plain text on solid backgrounds.
• Similar-colour merging so visually-identical foreground/background pairs roll up into one row with their worst-case ratio.
• Streaming results — each image's card appears as it finishes; no waiting for the batch.
• Exports:
  • PDF via pdfmake — selectable text, per-image page breaks, embedded swatches, simulations and clips.
  • Single-file Markdown with base64-embedded preview, simulations, swatches, and cropped regions. Best viewed in VS Code / Obsidian / Typora (GitHub strips embedded images from Markdown).

Quick start

npm install
npm run dev       # local dev server (http://localhost:5173)
npm run build     # produces dist/ — deploy anywhere as static files
npm run preview   # preview the built dist/ locally

The postinstall / prebuild / predev hooks copy the PaddleOCR ONNX models and the onnxruntime-web WASM binaries out of node_modules and into public/ so they're served as static assets.

How it works

File -> decode + downscale to 800-1400 px long edge (sRGB, no colour management)
     -> canvas + ImageData
     -> PaddleOCR PP-OCRv4 detection + recognition (via onnxruntime-web)
     -> per-detection vertical-strip scan
        -> k-means (k=2) on per-pixel luminance
        -> minority cluster = text, majority = background
        -> WCAG contrast ratio
     -> merge similar-colour findings
     -> per pair: WCAG AA/AAA, APCA, CVD contrast, vestibular, cognitive
        -> rolled-up overall verdict (PASS / WARN / FAIL)
     -> verdict (PASS / FAIL / NO_TEXT)
     -> render summary table + per-image cards (expandable rows
        + colour-blindness simulations)
     -> optional PDF / Markdown export

WCAG thresholds applied:

• AA: 4.5:1 normal text · 3:1 large text
• AAA: 7:1 normal text · 4.5:1 large text
• Large text = OCR bbox height ≥ 24 px (in the canonical resized image space).

See src/core/contrast.js for the contrast-ratio math.

Browser support

Browser	Status	Notes
Chrome / Edge (desktop)	Full	WebGPU OCR acceleration
Safari (desktop)	Full	WASM OCR (5-15s per image)
Firefox (desktop)	Full	WASM OCR
Chrome (Android)	Full	WASM OCR
Safari (iOS / iPadOS)	Full	WASM OCR; WebGPU disabled (ORT's WebGPU backend is unsupported on iOS)

First use downloads about 44 MB total (PaddleOCR detection + recognition models + ORT WASM). A service worker caches them in the Cache Storage API, so later visits load them without re-downloading.

Privacy

All processing happens in your browser. No images, OCR text, or contrast data ever leaves your device. Network traffic is limited to the one-time fetch of the PaddleOCR model files and the onnxruntime-web WASM binaries (served from the same origin as the app), plus a cookieless, privacy-friendly GoatCounter page-view count — no cookies, no personal data, no cross-site tracking.

Full details: Privacy Statement

Tech stack

• Vite — build tool
• Vanilla ES modules — no framework
• @gutenye/ocr-browser — PaddleOCR-in-the-browser
• onnxruntime-web — ONNX model inference
• APCA — perceptual-contrast math, vendored into src/core/apca.js (not an npm dependency)
• pdfmake — PDF generation

Hosting

The tool is hosted on GitHub Pages and deployed automatically from the main branch via .github/workflows/deploy.yml. The canonical live URL is:

image-colour-contrast-checker.timdixon.net

The same build is also available at the underlying Pages URL: timdixon82.github.io/Image-Colour-Contrast-Checker.

GitHub Pages cannot send custom HTTP response headers. The COOP/COEP headers needed for multi-threaded WebAssembly are instead injected by the service worker at public/sw.js. See docs/patterns/cross-origin-isolation-service-worker.md for details.

Dependencies and licensing

Every runtime and build-time dependency uses a permissive licence (MIT / Apache-2.0 / BSD-3-Clause / ISC / Boost-1.0). The one piece of non-permissive code is the APCA contrast math, which is vendored into the source tree rather than installed as a dependency — see A note on the APCA licence below.

The GoatCounter analytics client (public/count.js, ISC) is also vendored — self-hosted rather than loaded from gc.zgo.at so it works under the app's cross-origin isolation (see Privacy).

Direct dependencies

Declared in package.json — imported directly by the source, or consumed directly by the build:

Package	Spec	Licence	Role
@gutenye/ocr-browser	^1.4.8	MIT	Browser PaddleOCR PP-OCRv4 pipeline
onnxruntime-web	^1.26.0	MIT	ONNX inference runtime — imported by src/adapters/paddle-ocr.js
pdfmake	^0.2.10	MIT	PDF export
@gutenye/ocr-models	^1.4.2	MIT	Packaged PP-OCRv4 ONNX models — copied into public/models/ by copy-models.mjs (dev dependency)
vite	^5.4.10	MIT	Build tool (dev dependency)

onnxruntime-web is declared explicitly even though @gutenye/ocr-browser also depends on it: src/adapters/paddle-ocr.js imports it directly, so it is a first-class dependency rather than an implicit transitive one. @gutenye/ocr-models is declared for the same reason — scripts/copy-models.mjs reads its files at build time.

A note on the APCA licence

The APCA contrast math in src/core/apca.js is vendored verbatim from apca-w3 0.1.9 — it is not an npm dependency. Vendoring is deliberate: the apca-w3 package depends on colorparsley (AGPL v3), a CSS-colour-string parser this project never needs because colours are always parsed from hex by src/core/contrast.js. Copying only the two W3-licensed functions the tool calls (APCAcontrast and sRGBtoY) keeps APCA working while leaving the AGPL code out of the dependency tree entirely.

APCA is not permissively licensed. It is distributed under a "Limited W3 License": the W3C Software and Document License, narrowed by the APCA author to permit use only for predicting the contrast of web-based content in support of WCAG accessibility guidelines. Commercial use beyond that scope requires a separate written agreement, and uses unrelated to web-content contrast (medical, clinical, safety-related, aerospace, transportation, military) are prohibited.

This tool uses APCA solely to predict the contrast of text in web-UI screenshots for accessibility review — squarely within the licence's permitted scope. The vendored file keeps its original W3 licence header and unmodified G-4g constants. If you fork this project, be aware that src/core/apca.js is not covered by this repository's MIT licence and carries the usage restrictions above. APCA is also still in public beta and is not a WCAG requirement; it is reported here as an additional perceptual signal only.

Notable transitive dependencies

Package	Licence	Pulled in by
@gutenye/ocr-common	MIT	@gutenye/ocr-browser
@techstark/opencv-js	Apache-2.0	@gutenye/ocr-browser — contour finding
onnxruntime-common	MIT	onnxruntime-web
tiny-invariant	MIT	@gutenye/ocr-browser
js-clipper	Boost-1.0	@gutenye/ocr-browser
protobufjs + @protobufjs/*	BSD-3-Clause	onnxruntime-web
flatbuffers, long	Apache-2.0	onnxruntime-web
pdfmake's @foliojs-fork/* family (pdfkit, fontkit, linebreak, restructure)	MIT	pdfmake
Rollup, esbuild, PostCSS	MIT	vite

The complete, version-pinned tree is in package-lock.json.

Models

The OCR models are PaddleOCR PP-OCRv4 (Apache-2.0), distributed by @gutenye/ocr-models and copied into public/models/ at install time:

File	Source	Licence
ch_PP-OCRv4_det_infer.onnx	PaddleOCR	Apache-2.0
ch_PP-OCRv4_rec_infer.onnx	PaddleOCR	Apache-2.0
ppocr_keys_v1.txt	PaddleOCR	Apache-2.0

Generated PDF fonts

The exported PDFs are typeset in Roboto, bundled inside pdfmake's vfs_fonts.js. Roboto is licenced under Apache-2.0 (upstream).

Licence

This project is released under the MIT Licence — see LICENSE.

MIT was chosen because every runtime and build-time dependency is permissively licenced (MIT, Apache-2.0, BSD-3-Clause, ISC, BlueOak-1.0.0, Boost-1.0, MIT/Zlib), and MIT is fully compatible with all of them. The one exception is the vendored APCA math in src/core/apca.js (Limited W3 License) — this repository's MIT licence covers the project's own code, not that vendored file, which keeps its own usage restrictions; see A note on the APCA licence.

If you redistribute this project, please retain attribution for the upstream projects listed above — in particular PaddleOCR (Apache-2.0) for the OCR models and OpenCV (Apache-2.0) for @techstark/opencv-js. Apache-2.0 requires that the licence text and any NOTICE files travel with redistributed binaries. Both projects' licence files ship inside their node_modules directories and inside the published dist/ build.

Acknowledgements

• PaddleOCR — the upstream OCR project from PaddlePaddle.
• gutenye/ocr — the browser-first port of PaddleOCR that makes this app possible.
• ONNX Runtime — Microsoft's cross-platform inference engine.
• OpenCV.js (TechStark) — OpenCV in the browser.
• pdfmake — client-side PDF generation.
• APCA — Andrew Somers / Myndex; the Advanced Perceptual Contrast Algorithm.
• WCAG 2.2 — the W3C guideline whose contrast thresholds are encoded here.