Turns a scanned, image-only PDF into an editable Word document: OCR'd prose with headings, bullets and inline bold, ruled tables rebuilt as real Word tables at their original column proportions, multi-column pages read in reading order, figures embedded, and low-confidence words highlighted in place. Every table is also written to a parallel spreadsheet. Every run writes a quality report naming what it repaired and what a human still has to check.
Pages are analysed in parallel and cached, so a 40-page scan takes roughly twenty seconds cold and a few seconds warm.
Built and tuned against real-world scanned tenders and reports: mixed letterheads and stamps, ruled tables, multi-column layouts, and pages transposed during scanning. It is not a general document-layout engine; see MANUAL.md for what it deliberately does and does not handle.
Most OCR tools stop at recognised text. This one also rebuilds structure: table geometry, reading order, emphasis, and it tells you exactly what to check rather than presenting a wall of text as fact. The reasoning behind each design choice, its architecture, and its limits: MANUAL.md.
Requires Python 3.12 or later (set by scipy's own floor). Tested on 3.14.
Once per machine, install Tesseract, then the Python dependencies.
| Platform | Install Tesseract with |
|---|---|
| Windows | winget install --id UB-Mannheim.TesseractOCR -e |
| macOS | brew install tesseract |
| Linux | apt install tesseract-ocr (Debian, Ubuntu) or dnf install tesseract (Fedora) |
pip install -r requirements.txt
The Python code is identical across platforms. Tesseract is found on
PATH, at the default install location, or via --tesseract. Full
setup detail, including what each dependency does: MANUAL.md.
python scan_to_docx.py "path/to/scan.pdf"
No arguments launches an interactive menu instead. pdfocr.bat (Windows)
and pdfocr.sh (macOS, Linux) are thin wrappers that forward all arguments
and the exit code, so pdfocr scan.pdf --exclude-pages 1 works the same as
calling the script directly.
Every flag, day-to-day usage, reading the run output, and troubleshooting: MANUAL.md.
MIT. See LICENSE.
If this saved you time, buy me a coffee.