What is a Image to Text (OCR)?
Optical character recognition — OCR — is the process of turning a picture of text back into text. The image is a grid of coloured pixels; the output is characters in an encoding. Between the two sits a model that has been trained on millions of labelled lines and has learned which arrangements of dark pixels correspond to the letter e. This tool runs that model in your browser, so the image is read on your own machine rather than on a server.
OCR quality is decided mostly before the model runs. Text below roughly twenty pixels of height gives it too little information, which is the usual reason phone screenshots of documents and photographed receipts fail. A shadow across the paper is read as a dark band the model tries to interpret. Slight blur merges e, c and o into each other. The controls here exist for those three problems: enlarge the image, choose the one language that is actually present, and let the pre-processing throw away a background if the source is a clean scan.
The confidence figure the engine reports is the average per-character certainty of its own guesses. It is useful as a relative signal — a clean page comes back in the nineties and a bad photograph in the sixties — but it is not an accuracy measure. A model can be confidently wrong, most often on a character that looks like another, or on a word that is not in the language dictionary it was given.
How to extract text from an image
- Choose a screenshot, a scan or a photograph of a page. Nothing is uploaded; the file is handed to a Web Worker in this tab.
- Check the preview. If the text is not legible to you at normal size, it will not be legible to the engine either — scale it up before running.
- Set the language. Choose Bengali, or Chinese, and nothing else where you can; a second language widens the dictionary the engine uses to resolve ambiguous shapes.
- Pick a scale. 1.5× or 2× is the single most reliable fix for small print, at the cost of speed.
- Choose pre-processing. Greyscale always helps slightly; thresholding to pure black and white helps a clean scan and hurts a photograph.
- Press Read the text. The first run downloads several megabytes of engine and language data, which the browser then caches.
- Correct anything the engine got wrong in the output box, then copy it or download it as a .txt file.
Image to Text (OCR) limits and rules
| Item | Limit |
|---|---|
| Languages offered | English, Bengali, Hindi, Arabic, Chinese (simplified), Spanish, French and German |
| Default download | About 5–8 MB the first time, cached afterwards; a second language roughly doubles it |
| Scales | Original, 1.5× or 2× before recognition |
| Pre-processing | Greyscale and an optional average-luminance threshold |
| Handwriting | Not supported — the model is trained on printed text and does poorly on cursive |
| Output | Editable plain text, copyable or downloadable as a .txt file |
| Confidence | The engine's average per-character certainty, not an accuracy score |
| Uploading | None — the engine runs in a Web Worker and the image is read in the tab |
| Large images | Slower rather than blocked; 2× on a phone photograph can take a minute or more |
When to use it
- Getting the text out of a screenshot you can no longer find the source for.
- Copying a printed paragraph into a document without retyping it.
- Reading an old scan where the original file has been lost.
- Extracting Bengali text from a photographed page when English-only tools mangle it.
- Pulling the numbers off a receipt or a meter reading into a spreadsheet.
- Photos of confidential documents — a passport, a bank statement — where uploading is not acceptable.
Is it private?
The image is never uploaded. The recognition engine and its language data are downloaded from a public CDN, and the image is handed to a Web Worker running inside this tab and read from local memory. No pixel of your picture reaches this site or anybody else. That matters most for exactly the documents people are least willing to upload — identity papers, medical letters, bank statements — and it is the reason the first run is slow: the engine has to arrive before it can work.
Frequently asked questions
No. The OCR engine is a WebAssembly build that the browser downloads and runs locally, and the image is passed to a Web Worker in this tab. The only network requests are for the engine and its language data, which are identical for every visitor and contain none of your data. You can confirm it with the browser's network inspector: after the first run the tool works with the connection disconnected.
Because the engine and a language model — about 5 to 8 MB, and rather more for some scripts — have to be downloaded before any recognition can happen. The browser caches them, so the second run starts immediately. If you choose two languages at once, expect roughly double the download and noticeably slower recognition, because the engine has to consider both dictionaries for every ambiguous character.
Scale the image up first; text below about twenty pixels of height is the most common cause of failure and 2× fixes many cases outright. Then check the language — an engine looking for English will correct Bengali shapes into nonsense with perfect confidence. Greyscale almost always helps slightly, thresholding to black and white helps a clean scan and hurts a photograph. If the source is blurry or poorly lit, no setting will fully recover it.
No, and neither do the free alternatives. The engine is trained on printed type, and handwriting varies too much between people for this class of model. Neat block capitals on a form sometimes come through; joined-up writing almost never does. Some paid services specialise in handwriting, but even those expect you to correct the result by hand.
It is the engine's average certainty about each character it output, on a scale from 0 to 100. Treat it as relative rather than absolute: a clean page scores in the nineties and a blurry photograph in the sixties, and that difference is meaningful. It is not an accuracy score — the engine can be confidently wrong about a character that closely resembles another, so always read the result before trusting it.
Not directly. This tool takes images. If a PDF contains a scan, convert its pages to images first — the PDF to Image tool on this site will do that in your browser — and read each page. If the PDF already contains selectable text, there is no need for OCR at all: copying the text out of any PDF reader will be faster and more accurate.
Related tools
Written and maintained by Md. Nur Hossain Repon.