Shuna Maekawa / Projects

OCR-VFS — Visual Fidelity Scorer for OCR-to-Word Pipelines

2026-05-09 · Python, OCR, Azure Document Intelligence, Claude, OOXML

A research PoC that scores how well OCR-to-Word pipelines preserve the layout of scanned documents, and rebuilds layout-faithful DOCX files.

Overview

OCR engines are judged on text accuracy (CER/WER), but for old printed documents the text is usually fine — what breaks is the layout: headings, paragraphs, tables and where things sit on the page. OCR-VFS measures that. It renders a pipeline's Word output back to an image and scores it against the original scan.

How it scores

What I found

Engineering