Chinese OCR — image to text

The picture

Drop a photo here

or click to browse — you can paste a screenshot too.

Text

The first run downloads PaddleOCR PP-OCRv5, about 15–20 MB, Apache-2.0. After that it stays in this browser.

Printed Chinese and receipts are the point. Handwriting and stylised lettering are often wrong — that is a limit of the model, not a setting you missed.

Acceleration: Checking what this browser can do…

What leaves this tab

The picture does not. Recognition runs in this tab. There is no upload endpoint.

Model files are fetched from the PaddleOCR CDN when you press the button — never your photo.

About this tool

English OCR is a red ocean. Chinese on a phone photo is the gap

PaddleOCR.js runs PP-OCRv5 with ONNX Runtime in this tab. Detection and recognition are the official Apache-2.0 pipeline, about 15–20 MB the first time, then cached. Traditional Chinese and thermal receipts are the reason this page exists: Tesseract's Chinese accuracy is the thing people bounce off, and we are not pretending otherwise.

Drop, paste, or try the sample

The same paste path as Redact Screenshot — a screenshot on the clipboard is enough. The sample button draws a fake receipt so you can see the round-trip before you trust a real one. Output is plain text you can copy or download, plus a simple line list with scores so a weak box does not hide in an average.

Handwriting and display type will be wrong

That is the model, not a missing checkbox. Printed menus, invoices, and a phone photo of a page of 繁體 are in-distribution. A signature, a neon sign, or a calligraphy scroll is not. The page says so next to the button rather than in a buried FAQ.

Later, PDF

The 1201 PDF line still needs an OCR fallback for scanned pages. When /pdf-to-text ships, it will call this engine rather than a second one. Redact Screenshot is the sibling for when you have read the text and now need to hide it.

Frequently asked questions

Is the image uploaded?

No. Recognition runs in this tab. Model files are fetched when you press the button; the photo is not part of that request.

Which languages does it actually read?

The PP-OCRv5 ch preset is built for Chinese and still reads embedded English (receipts mix both). It is not a fifty-language pack. If your page is Japanese-only or Latin-only, expect a worse result than a specialist model.

Why not Tesseract?

Tesseract in the browser is fine for English. On Traditional Chinese from a phone, PP-OCRv5 is the reason to bother shipping a 20 MB model. That is the differentiation, stated plainly.

Can it read handwriting?

Poorly. Treat a bad handwritten result as the expected outcome, not a bug to file. Printed text is the supported path.

Will this become the PDF OCR fallback?

That is the plan. /pdf-to-text is on the 1201 line; when scanned pages need OCR, they should call this engine instead of growing a second one.