Image to Text

How it works

1. 1Upload a PNG, JPEG, WebP, TIFF or PDF file.
2. 2Tick the language or languages in your picture: English, Spanish or Japanese, up to three.
3. 3Choose "What the picture holds": Document or screenshot, One block of text, or Scattered text.
4. 4Click "Read the text" and wait for the text to appear in the result box.
5. 5Copy the text, or download it as .txt, .tsv or .hocr.

Example output

Test picture (generated), top part

Text read, default settings, 92 words

How to turn an image into text

Start by uploading your file, then pick the language and the layout that match the picture. The tool reads the picture on our server and shows the text in a box in the result panel. You can copy the text straight away or download it as a text file. To convert image to text this way, the letters should be sharp, dark on a plain background, and not too small. The picture to text workflow works best with clear, printed letters.

If the first result looks wrong, change the layout setting and run the job again. Compare the output with the original before you use it, because the engine can misread a letter or drop a line. Each job handles one file, so repeat the steps for each picture.

What this image to text tool does (and what it is not)

This tool takes one file at a time, either a picture or a PDF, and returns the words it finds as plain text, TSV or hOCR. It is text recognition only. It gives you the letters it reads, but it does not keep the layout of the page, and it does not keep rows and columns.

Photo to text results depend on how the picture was taken. Treat the output as a draft to check, not as a finished copy. The limits that apply to your use are in the tool's notes and below.

Which pictures read well: screenshots, scans, PDFs

Clean screenshots, scans and clear pictures of printed text read well. In our tests, clean screenshots had an error rate of 0.7%, and simulated scans at 150-200 dpi, tilted slightly and with noise, had 0.0%. These are measurements on generated test files, not on real documents, so results vary.

• Clean screenshots: 0.7% errors
• Simulated scans at 150-200 dpi: 0.0% errors
• A scan at 100 dpi: 1.0% errors
• Spanish text with accents and ñ: 0.0% errors

Which pictures do not: handwriting, photos, tables and receipts

Handwriting and artistic fonts are not something this engine is built for, and we did not test handwriting, so expect poor results. Vertical Japanese is not supported. Photos are harder. Our simulated phone photos of a page drawn with perspective had 44.8% errors on the default setting and 5.3% with One block of text. We did not test real phone photos, so we cannot say how they will do.

Tables and receipts need care. On our test table, the header row was lost and rows came out in the wrong order, with 35.0% errors on average over the table and receipt tests. A receipt with One block of text read at 0.7%, but the output is still a stream of text.

• Handwriting: not tested, expect poor results
• Vertical Japanese: not supported
• Simulated phone photos: 44.8% errors on default, 5.3% with One block of text

Measured accuracy on our test pictures

We measured accuracy on 20 pictures that we generated for this project. They include web and app screenshots, simulated scans, tilted pages, columns, a table, a receipt, Spanish and Japanese text, and two simulated phone photos. Accuracy is the share of characters that differ from the known true text, with default settings. Your own files may give different results.

• A dark terminal screenshot was inverted automatically and read with 0% errors.
• A screenshot only 500 px wide had 6.9% errors, and low-resolution pictures averaged 3.9%.
• A page turned 90 degrees read with 0.0% errors.
• A page tilted 7 degrees returned nothing at first, then read with 0.0% after the automatic retry.
• Two and three column pages read with 0.0% errors and the correct order on the default setting.
• Japanese horizontal text had 2.5% errors, mostly in individual kanji.

What the reader's own score means

The reader's own score is the mean confidence from 0 to 100. It shows how sure the engine was of each letter. It is not a grade of the whole result, and it does not show missing lines or wrong order. In one of our tests, a table scored 96 and still lost rows and mixed their order.

Our demo is a 1000 x 640 px article screenshot with the default settings. It returned 92 words with a score of 94.8. The text was right except that two bullet marks read as + and *, and the white-on-blue button label "Read more articles" came out as "ad more articles". Check labels on colored buttons before you use them.

Image to text languages: English, Spanish and Japanese

The tool has English, Spanish and Japanese (horizontal text) installed. You can tick up to three languages together, for example English and Spanish for a bilingual page. Other languages are not supported, and they come out as wrong letters. Vertical Japanese is not supported at all.

Spanish with accents, ñ and ¿ read with 0.0% errors in our tests. Japanese horizontal text had 2.5% errors, with individual kanji wrong, so check any Japanese result closely.

Pick the right setting: document, one block, scattered text

The default, "Document or screenshot", suits most pages, including pages with columns. Use "One block of text" for a receipt or a photographed page that is one block of text. Use "Scattered text" for labels in a diagram, where words sit apart from each other.

The setting makes a large difference. On our two-column tests the default kept the reading order, but One block of text put the order out, with 66.6% errors. On simulated phone photos, One block of text cut errors from 44.8% to 5.3%.

• Document or screenshot (default): pages, screenshots and columns
• One block of text: receipts and photographed pages
• Scattered text: labels in a diagram

Tilted, dark or small pictures: what the tool fixes on its own

The tool handles several common problems automatically. It turns a picture upright using its EXIF orientation and flattens transparency to white. A picture with a dark background and light text is inverted automatically. A small picture, with its long side under 1,600 px and low confidence, is enlarged and read again, and the better result is kept. Pictures over 10 megapixels are scaled down.

If the first read finds no text at all, the engine tries once more with straightening on. You can also tick "Straighten tilted text first (up to 15 degrees)" yourself. Automatic fixes help, but they do not guarantee a good result for every file.

Read a scanned PDF to text (first pages only)

You can read a scanned PDF. Without an account the tool reads the first 2 pages, and signed-in users can read up to 5 pages. Each page is turned into a picture and read as a picture, so any text layer already in the PDF is not used. The page count you choose sets how many pages are read.

In the result, pages are separated by a line holding only a form-feed character. If the PDF has more pages than were read, the result panel says so. Signed-in users pay 1 credit for each page read.

Text, TSV or hOCR output

Plain text is saved as .txt in UTF-8. TSV gives each word with its position and confidence, which helps if you want to check words one by one. hOCR gives position markup and is delivered as plain text. Choose the format before you run the job.

The result panel shows the text in a box, with a Copy button and a Download button. The download is named text.txt, or .tsv or .hocr for those formats. If you want a plain image to txt file, choose the .txt format. The tool does not make a searchable PDF, so the output is text, TSV or hOCR only.

If no text is found

Sometimes the picture holds no text the engine can read. The job still succeeds, with an empty result. The panel then says "No text was found", and no credits are charged.

Try the "One block of text" setting, tick "Straighten tilted text first", or use a larger, sharper picture with dark text on a plain background. Failed and timed-out jobs are refunded in full, so a job that does not finish costs you nothing.

Read text from an image in code

You can use the API to extract text from image files in your own code. Send the raw file bytes in the body of a POST request to https://api.aiipworld.com/v1/image-jobs?op=image_to_text. The Content-Type must be image/png, image/jpeg, image/webp, image/tiff or application/pdf. Send your key as Authorization: Bearer <API key> or as x-api-key.

Settings go in the query string: lang, psm, format, dpi, max_pages, deskew, auto_rotate, invert, upscale and timeout_sec. Unknown names are rejected. The response is 202 with a job_id. Poll GET /v1/jobs/:id until the job succeeds, fails or times out, then download the text from the short-lived download_url. The page shows a ready cURL, Node and Python example built from the settings you chose. For the general API, see /docs/.

Possible errors are 400 invalid_params, 413 file_too_large, 422 not_a_pdf, 402 insufficient_credits and 429 rate limited. A failed job reports an error_code.

Privacy: what happens to your picture

Your file is sent to our server, and nothing is processed in your browser. The file is used only to make your result. The working copy of your upload is deleted as soon as the job ends, whether it succeeds or fails.

In the free web tool, the result is deleted after 10 minutes, and no history is kept. Download links are short-lived. Anonymous use needs a human check by Cloudflare Turnstile. You must have the right to use what you upload, as set out in the Acceptable Use Policy.

Questions