PDF to DOCX local OCR checks, 2026-10-16
Three self-made image-only US Letter PDFs; 1836x2376 px source images.
English: Arial 44 px. Chinese: Microsoft YaHei 44 px. Table: Arial 44 px.
Each sample: 1 source page. Language eng / chi_sim+eng / eng respectively.
Mode layout; local existing Tesseract assets, PSM 3, DPI hint 216.
Render min(3,2400/max(page width,page height)); Letter renders at scale 3.
Pre-OCR build rejects image-only pages. Current OCR prototype initially broke
English line order and Chinese spacing after recognition; current reconstruction
uses OCR line IDs for geometry and recognized full lines for single-column pages.
Table: native 4x3; source PEN001/BOOK002/TOTAL003 recognized as
PENO001/BOOKO002/TOTALO003. No spelling/identifier correction was applied.
English and Chinese final paragraph strings exactly match expected.json.
Actual browser downloads parsed with python-docx; LibreOffice 26.8.1 independently
opened and rendered all 4 OCR outputs, 1 page each. Microsoft Word not installed/tested.
Six original text-layer outputs: all unzipped package parts match old baselines.
Existing 2026-10-13 kit remains historical; it documents pre-OCR behavior.
No physical-phone timing, cold-network timing or general accuracy measured.
