Image to Text OCR – Extract Text from Photos, Scans & Screenshots

Extract text from images in bulk with the open-source Tesseract OCR engine. Give it image URLs and get clean text back, with confidence scores, optional line and word bounding boxes, and automatic rotation fixing. OCR runs inside the Actor, so images are never sent to a third-party API.

• 🖼️ Formats: JPEG, PNG, WEBP, multi-page TIFF, BMP, GIF, HEIC/HEIF (iPhone photos), AVIF.
• 🌍 27 languages, mixable (up to 4 at once): English, German, French, Spanish, Italian, Portuguese, Dutch, Polish, Czech, Swedish, Danish, Norwegian, Finnish, Turkish, Russian, Ukrainian, Greek, Vietnamese, Indonesian, Japanese, Chinese (Simplified/Traditional), Korean, Arabic, Hindi, Thai, Hebrew.
• 🔄 Auto-rotation of sideways or upside-down scans, EXIF orientation for phone photos, automatic upscaling of small screenshots.
• 📐 Layout modes for documents, single text blocks (receipts, labels) and sparse text (signs, posters, product photos).
• 🤖 Built for automation and AI agents. Structured JSON output, per-image error codes, and you only pay for images that were processed.
• 💸 $0.003 per image ($3 per 1,000).

Use cases

• Digitise scanned letters, forms, archives and typewritten documents.
• Pull text out of screenshots, slides, memes and social-media images for search, moderation or RAG.
• Read product labels, packaging, shelf tags and street signs from photos.
• Pre-process receipts and invoices before parsing with an LLM.
• Make image libraries searchable (alt-text and indexing).

Input

{
  "imageUrls": [
    "https://upload.wikimedia.org/wikipedia/commons/e/e2/Typewritten_letter_from_Dr_Claude_H_Organ_Jr_MD_FACS_to_Dr_Leonidas_Berry_dated_October_5_1965.jpg",
    "https://upload.wikimedia.org/wikipedia/commons/a/a5/Street_sign%2C_Primrose_Street_-_geograph.org.uk_-_6556772.jpg"
  ],
  "languages": ["eng"],
  "pageMode": "auto"
}

Field	Default	Description
imageUrls	–	Image URLs (required), up to 10,000 per run
languages	["eng"]	Up to 4 language codes (eng, deu, jpn, chi_sim... or en, de, ja)
pageMode	auto	auto, single-block (receipts/labels), sparse (signs/photos)
autoRotate	true	Fix sideways/upside-down text
includeLines	false	Lines with bounding boxes and confidence
includeWords	false	Words with bounding boxes and confidence
maxPagesPerFile	20	Pages read from multi-page TIFFs
maxConcurrency	2	Images processed in parallel

Output

{
  "url": "https://upload.wikimedia.org/.../Street_sign%2C_Primrose_Street_-_geograph.org.uk_-_6556772.jpg",
  "status": "ok",
  "format": "jpeg",
  "width": 640,
  "height": 480,
  "pageCount": 1,
  "pagesProcessed": 1,
  "languages": "eng",
  "confidence": 92.7,
  "rotated": 0,
  "wordCount": 3,
  "text": "PRIMROSE\n\nSTREET EC2",
  "warnings": []
}

• Multi-page TIFFs also get a pages array.
• includeLines/includeWords add lines/words with box: [x, y, width, height].
• Errors (HTTP_404, NOT_AN_IMAGE, UNSUPPORTED_FORMAT, TOO_LARGE, CORRUPT_FILE...) are reported per image and not charged.
• PDFs are rejected with a pointer to our PDF to Markdown Actor.

Pricing

Pay per event, plus a tiny Actor start fee ($0.00005 per GB of memory):

• Image processed: $0.003 per image (each page of a multi-page TIFF counts as one image). That's $3 per 1,000 images, with compute included.
• Failed downloads, non-images, PDFs and oversized files are free.
• Images with no recognisable text are charged, because the OCR work was done (they return text: "" and a warning).

Popular OCR Actors on the Store charge around $5 per 1,000 images.

Accuracy (honest)

• Tesseract is excellent on printed text: documents, screenshots, typewritten pages, signs with clear fonts (typically 85–95% confidence).
• It is weak on handwriting, very stylised fonts, heavy perspective, low-light photos and text over busy backgrounds. Check the confidence field: below ~60 usually means poor results.
• For best results use images at least ~1,000 px wide, choose the correct language(s), and use sparse mode for photos with scattered text.
• Tables come out as text lines (no table structure). For layout-heavy PDFs use a PDF converter instead.

FAQ

Are my images stored or sent elsewhere? No. They are downloaded into the run container, OCR'd locally and discarded. Only the extracted text is saved to your dataset.

Can I OCR images behind a login? No. URLs must be publicly downloadable (or pre-signed links such as S3/GCS signed URLs).

How fast is it? About 1–3 seconds per typical image per CPU core. Raise memory (4 GB = 1 core) and maxConcurrency together for big batches.