OCR Japanese, Chinese, Korean & English — Images & Scanned PDFs

Extract text from photos, screenshots and scanned PDFs in Japanese, Chinese (Simplified / Traditional), Korean and English — with output that is ready for search, translation or an LLM.

• No stray spaces between CJK characters (a common problem with OCR engines): 東 京 都 → 東京都
• Vertical Japanese (縦書き) supported
• Smart PDF handling: pages that already contain text use the embedded text (exact and cheap); only scanned pages are OCR'd
• Noise filtering: photo edges, shadows and stamps that OCR engines misread as garbage are dropped
• Frames and table lines removed before OCR: characters next to borders, boxes or big brackets are not lost (令和 stays 令和, not 和)
• Pay per page — failed downloads and unsupported files are never charged

What can it be used for?

• Digitize Japanese / Chinese / Korean documents: invoices, receipts, notices, contracts, forms
• Make scanned PDFs searchable or feed them to an LLM / RAG pipeline
• Extract text from screenshots and product photos for translation
• Process vertical Japanese text from books, newspapers and signs
• Call it from AI agents via the Apify API or MCP

Accuracy (our own measurements)

Character error rate (CER, lower is better), measured on our test set — printed text, 300 DPI:

Test	CER
Japanese government document (4 pages incl. a framed date box, rendered at 300 DPI, compared with the original text)	1.8%
English research paper page (two columns)	0.3%
English paper title page (author grid, margin stamp)	0.6%
Simplified Chinese sample	0.0%
Traditional Chinese sample	2.8%
Korean sample	0.0%
Japanese "phone photo" sample (blurred, rotated, JPEG)	2.8%
Vertical Japanese sample	0.0%

CER counts misread characters; text that is correct but in a different reading order (for example an author grid read row by row) is not counted as an error. Multi-column grids such as author lists are read row by row. Results depend on image quality. Handwriting is not supported.

Input

Field	Description
fileUrls	Direct links to images (JPEG, PNG, WebP, TIFF, GIF, BMP, AVIF) or PDFs. Only these URLs are downloaded.
languages	Any of jpn, chi_sim, chi_tra, kor, eng (default: jpn + eng). Choose only what you need.
verticalText	Turn on for vertical Japanese
pdfTextLayer	auto (default): use embedded text when present, OCR only scanned pages. ocr: OCR every page.
pageRange	PDF pages, e.g. 1-5, 8
dpi	Render resolution for scanned PDF pages (default 300)
includeLines	Also return each line with confidence and bounding box
password	Password for encrypted PDFs
maxPagesPerFile, maxFileSizeMb, downloadTimeoutSecs, maxConcurrency	Limits

{
    "fileUrls": ["https://example.com/scan.pdf", "https://example.com/receipt.jpg"],
    "languages": ["jpn", "eng"]
}

Output

One dataset item per file:

{
    "url": "https://.../sample_notice_ja.jpg",
    "kind": "image",
    "status": "success",
    "languages": ["jpn", "eng"],
    "processedPages": 1,
    "ocrPages": 1,
    "textLayerPages": 0,
    "avgConfidence": 90.8,
    "charCount": 142,
    "text": "お知らせ\n来週の月曜日から、事務所の受付時間を変更します。\n\n新しい受付時間は、午前9時から午後5時30分までです。\n\nご不明な点は、総務部までお問い合わせください。\nNotice: Office hours change from next Monday (9:00-17:30).",
    "pages": [
        { "pageNumber": 1, "method": "ocr", "text": "...", "confidence": 90.8, "droppedNoiseLines": 1 }
    ],
    "warnings": []
}

status is success, no_text, partial (an error after some pages; those pages are included) or failed (with error). With includeLines, each page also has lines: [{ text, confidence, bbox: { x0, y0, x1, y1 } }].

Pricing

Pay per event — no subscription, platform usage included:

Event	Price
Page OCR (per image or scanned PDF page)	$0.008
Text-layer page (PDF page that already contains text — no OCR needed)	$0.0003
Actor start (per run, Apify standard)	$0.00005 per GB of memory

Failed downloads and unsupported files are not charged. You can cap the cost of a run with the maximum charge setting in Apify.

Tips

• Select only the languages in your documents — e.g. jpn alone for pure Japanese text.
• For small text in scanned PDFs keep dpi at 300.
• Low avgConfidence (below ~60) usually means a wrong language setting or a blurry image.

Limitations

• Handwriting is not supported.
• Tables are returned as text lines, not as structured tables.
• For PDFs that already contain text and need clean Markdown (headings, two-column papers, RAG chunks), see PDF to Markdown for LLM & RAG.
• Mixed horizontal and vertical text on the same page: choose the dominant direction with verticalText.

Privacy

Only the URLs you provide are downloaded. Files are processed in memory and the results are written only to your own run's dataset. Please make sure you have the right to process the documents you submit.