OCR a scanned Chinese PDF on your device

What this tool does, exactly

The reader on this page is PaddleOCR's — a model built in Chinese first — running in this browser tab rather than on somebody's server. A scanned contract, a form, a mixed Chinese and English page: every page is drawn and read here, and you get the text, the text as Markdown, and the scan made searchable, with the layer read back before the file is offered.

What was measured

A printed page of Simplified Chinese with English names and Arabic numerals mixed in, drawn at 200 dpi in a system font. This site's own test read 0.97 % of its characters wrong — under one in a hundred — with every line found, in about 1.4 seconds on one processor thread. Traditional Chinese, vertical text and handwriting were not measured and are not promised.

Which model, and why it is the default

PP-OCRv6 from PaddleOCR, a project whose home language is Chinese; one recogniser covers Chinese, Japanese and the Latin scripts, so a bilingual page needs no choice and no second download. Russian, Ukrainian and Belarusian are the one script that needs a second recogniser, and that is a choice in the options.

Mixed pages

A Chinese page with an English company name, a date in Arabic numerals or a product code is read as one page: the recogniser holds all three in one character set. Full-width punctuation — ，。：— is kept as it was printed.

The searchable PDF

The recognised characters are laid invisibly over the original pages and encoded so that a viewer's search finds them: the site's test reopens the file and looks for the words that were written. Copying from the searchable file gives the recognised text; the page image underneath is the untouched scan.

Where it stops

Vertical text is not laid out as columns; the reader finds horizontal lines. Handwritten characters are not read reliably. A coarse scan loses strokes, and a character with a lost stroke is read as its nearest neighbour — the doubtful-lines list under the result names the lines it was least sure of.