pdfmark-ocr 0.3.0

Fast, fully-local PDF & image (PNG/JPG/...) to Markdown converter with image OCR, plus a ZIP of parsed images. No API calls, runs 100% offline. Supports Python 3.9-3.12.

pdfmark-ocr

Fast, fully-local PDF → Markdown — with OCR for images. No API calls, no cloud, nothing leaves your machine.

Microsoft's markitdown only reads native text; if a PDF page is a scan or has text inside images, you lose it. pdfmark-ocr reads native text directly (instant) and runs OCR on embedded images, in parallel across your CPU cores.

Install

pip install pdfmark-ocr

That's it — works on Windows, macOS, and Linux (Python 3.9+). The default OCR engine is RapidOCR (ONNX, fast, small).

Usage

pdf2md document.pdf                       # -> full_stitched_output.md
pdf2md document.pdf -o notes.md           # choose output file
pdf2md document.pdf --workers 4           # control parallelism
pdf2md document.pdf --min-image 0         # OCR every image, even tiny ones
pdf2md scanned.pdf --page-ocr always      # render+OCR whole pages (scans/vector text)

You can also run it as a module:

python -m pdf2md_ocr document.pdf

Options

Flag	Default	Meaning
-o, --output	full_stitched_output.md	Output markdown path
--engine	rapidocr	rapidocr (default) or easyocr
--workers	auto	Parallel page workers
--max-dim	1500	Downscale images larger than this (px)
--min-image	16	Skip images smaller than this (px); 0 keeps all
--page-ocr	auto	Render+OCR whole pages for vector-drawn/scanned text. auto does only text-poor pages; always does every page; never disables it
--render-dpi	200	Resolution for full-page render OCR

Optional: EasyOCR engine

EasyOCR is heavier (pulls in PyTorch, hundreds of MB) but you may prefer its accuracy on some documents:

pip install "pdfmark-ocr[easyocr]"
pdf2md document.pdf --engine easyocr

How it works

1. Native text is extracted directly from the PDF — instant, no OCR.
2. Embedded images are downscaled and sent to the chosen OCR engine. Pages whose text is vector-drawn or scanned (no extractable native text) are rendered to a bitmap and OCR'd whole — controlled by --page-ocr.
3. Pages are processed in parallel, one OCR engine per worker process.
4. A per-worker cache means repeated logos/headers are OCR'd only once.

Use from Python

from pdf2md_ocr import stitch_full_pdf

stitch_full_pdf("document.pdf", "out.md", engine="rapidocr")

License

MIT