Manchu OCR Lite · Manchu.ai

01 Your page

Choose pagesJPG, PNG, TIFF, PDF, WebP, BMP, GIF … · multiple files

Drop images or PDFs anywhere on this page. Every PDF page and every TIFF frame becomes one page. Pages narrower than 1800 px are upscaled before detection, as in the SBB pipeline.

Sample pages

02 Crop & recognize

Word detector CRNN M20-e20 in both

Loading model…

Recognition is the mixed-corpus CRNN M20-e20 (SBB print + NGSYD handwriting). Use the bilingual detector on Manchu–Chinese pages such as 四書集註; it ignores Chinese text. Switching reloads the model; recognize the page again to compare.

Boxes with detector confidence ≥ 0.50 are kept; below 0.60 they are flagged, drawn dashed and shown in [ ] in the text (SBB delivery rule).

Small commentary script

Small script (double half-lines inside one column, as in 四書集註) is read as its own lines, left half then right. Auto turns this on when the page shows two clearly separate word sizes.

Reading order · advanced

Column model Column pitch Size threshold

Columns run left to right, words top to bottom; each small-script run splits into its left and right half-line.

kept ≥ 0.50flagged < 0.60small script

Run log

Reading layout Image label size

Choose a page to begin.

ᠮᠠᠨᠵᡠ

A page, then its words.

Open a page scan or pick a sample, then recognize.
Hover or click a recognized word for its QWJ definition.

Show annotations (boxes and labels)

QWJ dictionary 御製增訂清文鑑

Hover, focus, or click a word in the image or Roman text. Exact entries come first; lemma candidates and spelling suggestions are labelled. The dictionary is looked up inside this page when data/qwj.json is present; it is not part of the public release.

Definitions will appear here.