chrome-ocr

Turn local PDFs and images into Markdown with Chrome's hidden OCR engine.

No API key. No cloud bill. No model download. Chrome already ships the model; this project calls it directly.

• It reuses Chrome's built-in Screen AI component instead of shipping another OCR model.
• It stays local, so you can process documents without sending them to a cloud API.
• It returns layout-aware Markdown, not just a flat text dump.
• It handles the fast path for text-layer PDFs and only OCRs pages that actually need it.

This project is source-install first.

git clone https://github.com/ayismas/chrome-ocr.git
cd chrome-ocr
pip install -e .[pdf]

Check whether Chrome's OCR component is available:

chrome-ocr doctor
python -m chrome_ocr doctor

Extract a PDF to Markdown:

chrome-ocr pdf report.pdf -o report.md
chrome-ocr pdf report.pdf --pages 1,3,5-8 --dpi 300

OCR an image:

chrome-ocr img scan.png
chrome-ocr img scan.png -o scan.md

from chrome_ocr import ocr_img, ocr_pdf

markdown = ocr_pdf("report.pdf")
markdown = ocr_pdf("report.pdf", pages=[1, 3, 5], dpi=300)

text = ocr_img("scan.png")

Low-level engine reuse is also available:

from chrome_ocr import ScreenAIEngine, ocr_pdf

engine = ScreenAIEngine()
markdown = ocr_pdf("report.pdf", engine=engine)

• PDF to Markdown: direct extraction for text PDFs, OCR fallback for scanned pages.
• Image to Markdown: file paths, PIL.Image, and NumPy arrays are supported.
• Layout-aware output: headings, paragraph breaks, indented blocks, tables, and formulas are reconstructed from geometry.
• Cheap warm runs: the DLL is loaded once per process and reused afterwards.

Reports whether Chrome's chrome_screen_ai.dll is present and whether it initializes correctly.

OCR an image and print Markdown to stdout, or write it to a file with -o.

Extract or OCR a PDF and emit Markdown.

• --pages uses 1-based indexing.
• --page-sep controls how multiple pages are joined.
• --dpi affects scanned-page rasterization quality.

Returns only the content extracted from the PDF. It does not inject a file title or page headers.

OCR an image to layout-aware Markdown text.

ocr_img_md is an identical alias.

Advanced API for custom DLL paths and engine reuse across calls.

The repository includes a reproducible benchmark harness so you can generate numbers on your own machine instead of relying on screenshots.

python benchmarks/run_benchmark.py img path/to/scan.png --repeat 5
python benchmarks/run_benchmark.py pdf path/to/report.pdf --repeat 3 --pages 1-5

See benchmarks/README.md for methodology and external tool comparisons.

ocr_pdf() can still extract text-layer PDFs without Chrome Screen AI. The DLL is only required when a page has little or no embedded text and must be OCR'd.

Open Chrome, go to Settings -> Accessibility, enable any screen-reader-related feature, then confirm that this folder exists:

%LOCALAPPDATA%\Google\Chrome\User Data\screen_ai\

The Markdown formatter maps visual structure as follows:

Chrome ships an accessibility component called chrome_screen_ai.dll.

This library loads the DLL via ctypes, feeds it image data using the SkBitmap memory layout reverse-engineered from Chromium and Skia sources, and decodes the returned VisualAnnotation protobuf without compiling any .proto files.

The bounding-box metadata is then used to reconstruct document structure that most OCR wrappers throw away.

• benchmarks/README.md: reproducible benchmark workflow
• docs/launch-kit.md: maintainer launch copy and post templates
• screen_ai_pdf_parser.py: legacy compatibility wrapper for older imports

pip install -r requirements-dev.txt
pytest tests/ -v

Tests that require Chrome's DLL are skipped automatically when the component is not installed.

MIT. See LICENSE.

chrome_screen_ai.dll remains part of Google Chrome and subject to Google's terms. This project does not redistribute or modify the DLL.