Image to Text OCR: Chinese, Japanese & 50+ Languages

Extract text from images, photos, screenshots and scanned documents. Verified strong Chinese and Japanese recognition, plus 50+ other languages across Latin, Cyrillic, Arabic, Devanagari, Thai, Greek and more. Returns clean reading-order text, per-line confidence scores and bounding boxes.

Powered by RapidOCR (Apache-2.0), running Baidu's PaddleOCR PP-OCR detection and recognition models entirely inside the actor. There is no third-party OCR API in the loop, so results don't depend on another service staying up or rate-limiting you.

Great for: OCR API, extract text from image, image to text converter, Chinese OCR, Japanese OCR, reading signs and receipts, digitizing screenshots, scanned-document pipelines, and any workflow that needs text pulled out of pictures at scale.

What it does

Give it a list of image URLs, or upload an image directly. It runs OCR on each one and returns the text it found, with no files retained beyond the run.

Verified accuracy

Tested against real images before shipping, not just the library's own claims:

Test	Result
Clean screenshot (invoice-style text)	Exact match, ~99.7% avg. confidence
Lower-quality photo (rotated, blurred, noisy "sign")	Exact match despite the degradation
Real photo, dense Chinese e-commerce graphic	Exact match, ~97.8% avg. confidence
Real photo, Japanese train station sign (kanji + hiragana + romaji)	All real text read correctly
Japanese news headlines (dense text)	Exact match, ~99.9% avg. confidence
Korean news headline, language: korean	Correct, ~98% avg. confidence

Chinese and Japanese read correctly with the default auto setting: that's why they're in the title, not just a marketing claim. Korean needs the explicit language: korean option (auto-detect does not catch Korean reliably); once selected, it's accurate. Other scripts also have dedicated per-script models via language with generally good results, though accuracy varies more by script and image quality than the verified Chinese/Japanese/Korean paths above.

Example input

{
  "images": [
    "https://example.com/receipt.jpg",
    "https://example.com/screenshot.png"
  ],
  "language": "auto"
}

Input

Field	Description
images	List of image URLs (JPG, PNG, WEBP, BMP, TIFF)
uploadedImage	Upload a file directly instead of a URL
language	auto (default) handles Chinese, Japanese, English and most Latin-alphabet text. Set explicitly for Korean, Arabic, Cyrillic, Devanagari, Thai, Greek, Kannada, Tamil or Telugu if auto misreads a specific script.
maxImages	Safety cap on images processed per run (default 50)

Output

One row per image:

Field	Description
imageUrl	The image processed
text	Full extracted text, in reading order
blocks	Every detected text line: text, confidence (0-1), box (4-point bounding box)
blockCount	Number of text lines/blocks detected
averageConfidence	Mean confidence across all blocks
language	The language mode used
status	ok or error
error	Error message, if any

Pricing

Billed per image successfully processed. Failed images (bad URLs, non-image content, etc.) are not charged.

FAQ

How do I extract text from an image? Paste the image URL into images (or upload the file) and run. The text field in the output holds everything found, in reading order.

Does this work for Chinese and Japanese text? Yes. Verified against real Chinese and Japanese test images before shipping (see the accuracy table above). Leave language on auto.

Does this read Korean? Yes, but set language to korean explicitly; leaving it on auto will not read Korean reliably (verified). With korean selected, accuracy was ~98% in testing.

What about Arabic, Cyrillic, or other non-Latin scripts? Set language to the matching option. These use RapidOCR's dedicated per-script models; results are generally good but vary more than the verified Chinese/Japanese/Korean paths.

Does it work on photos, not just clean scans? Yes. Tested against a deliberately degraded (rotated, blurred, noisy) photo of a sign and it read every line correctly.

Do I get confidence scores and bounding boxes, or just plain text? Both. The text field gives you the plain reading-order output; blocks gives you each line individually with its confidence score and bounding-box coordinates, if you need to validate results or highlight regions.

Is this using Google Lens or another third-party OCR API? No. OCR runs entirely inside the actor via the open-source RapidOCR engine, no external API call, no dependency on a third-party service staying up.

Please note

Only process images you have the right to use. Images are processed transiently and not retained beyond the run.

Built with the Apify Python SDK + RapidOCR.