Can I extract text from a scanned PDF in Chinese, Japanese or Korean?
Chinese and Japanese worked in our test; Korean did not. On 2026-10-19 we ran six test pages through PDF to Text. The Chinese and Japanese pages came back with every character right, including a rotated, noisy one. The Korean pages came back as a few wrong characters, and the result carried a warning.
What did we test?
In our test we made six one-page PDFs that contain only a picture of text, so there is no text layer to copy. There are three languages with two versions each. The text is invented: a title, a few sentences about a quarterly report and a line of numbers such as 1,240 and 980. We drew it as an image at 200 dots per inch, in a standard Windows font for each language. The first version of each page is clean. The second is turned by 1.2 degrees, has about 60,000 specks of noise and is slightly blurred, to look a little more like a scan. We uploaded all six to PDF to Text on osmanthdocs.com and compared the text we got with the text we drew. The comparison ignores spaces and line breaks. The similarity score is the ratio that Python's difflib gives for the two texts: 1.0 means identical and 0 means nothing in common.
| Page | Characters drawn | Characters returned | Similarity | Label and note |
|---|---|---|---|---|
| Chinese, clean | 83 | 83 | 1.0 | "Looks good", noting that the PDF is scanned |
| Chinese, rotated and noisy | 83 | 83 | 1.0 | "Looks good" |
| Japanese, clean | 99 | 99 | 1.0 | "Looks good" |
| Japanese, rotated and noisy | 99 | 99 | 1.0 | "Looks good" |
| Korean, clean | 92 | 21 | 0.16 | "Review recommended", with "Low OCR confidence on page(s) 1. Check names and numbers." |
| Korean, rotated and noisy | 92 | 11 | 0.16 | "Some content may need correction", with "Very little content could be extracted from this PDF." |
The Korean text that did come back was not Korean. The clean page returned a few unrelated characters and part of the number line, such as 1,240 and 980, with wrong characters around them.
Why might Korean fail?
We did not look inside the model, so this is what the documentation suggests and not something we proved. The RapidOCR documentation lists several families of recognition models. It says its newest family covers Chinese, English, Japanese and many European languages, and lists Korean under older families. The PaddleOCR documentation for its multilingual model names Simplified Chinese, Traditional Chinese, English and Japanese as the default types, and describes a separate Korean model. Our server logs show the tool loading a small model of the newest family. Taken together, that fits what we saw, but the documentation does not say that this particular model cannot read Korean, and we cannot tell from it alone.
What should I check after converting?
- Read the label and the notes first. In our test, both Korean pages were flagged, so the failure was visible.
- Look at the numbers. Even the wrong Korean output kept some digits, so a line with digits is not proof that the rest is right.
- Compare the start and end of each paragraph with the scan.
- If the text is mostly Korean or another language we did not test, treat the result as unreliable until you have compared it with the page.
What are the limits?
This is six pages of computer-drawn text in one font per language, so it says nothing about handwriting, small print, photos of pages, creased paper, or vertical Japanese text. A real scanner or phone camera is harder than our "noisy" version. We tested three languages only, and our Chinese pages used simplified characters. For the general case of OCR on a scan, see how to extract text from a scanned PDF. For what the plain and cleaned text files contain, see why text copied from a PDF breaks at every line.
Your PDF is uploaded to our servers, processed with open-source software and deleted when processing finishes. The results stay for 60 minutes.