How to Extract Text From an Image (and When OCR Will Let You Down)

You have a photo of a textbook page, a screenshot of a slide, or a scanned handout, and you need the words as text you can edit.

That is OCR — optical character recognition — and for clean printed text it is close to instant. For everything else, the result depends almost entirely on the image you feed it. Most bad OCR output is a bad photo, not a bad tool.

The quick version

Our Image to Text tool runs OCR in your browser. Drop in a PNG, JPEG, or WebP up to 10 MB and the text comes back in a few seconds. Nothing is uploaded — the recognition happens on your own machine, which matters if the page is from something unpublished or confidential.

It reads English printed text. If the result looks like nonsense, the sections below are almost certainly why.

Give it a better image

OCR accuracy is decided before you ever open the tool.

Resolution is the big one. The engine needs enough pixels per character to tell an e from a c. As a rule, letters should be at least around 20 pixels tall. A screenshot of a full page at normal zoom does not clear that bar — zoom in before you capture, not after. For scans, 300 DPI is the standard target.

Crop to the text. Trim away the desk, the facing page, your thumb. Extra content gives the engine more chances to hallucinate characters out of noise.

Keep it straight. Skew is the second biggest accuracy killer. Photograph the page flat and square-on rather than at an angle, and straighten it before running OCR if it came out tilted.

Light it evenly. Shadows across the page, glare from a lamp, and the shadow of your own phone all hurt. Diffuse daylight beats a bright overhead bulb. Grayscale is fine — colour buys you nothing here.

Avoid re-compression. A screenshot pasted into a chat app, downloaded, and screenshotted again has lost detail at every step. Use the original capture.

Why the text comes back with strange line breaks

Run OCR on a paragraph and you often get output broken at every visual line, like this:

The integration of digital technologies into contemporary classrooms has transformed pedagogical approaches

That is not a bug. The engine reports what it sees, and what it sees is lines on a page. It has no way of knowing your paragraph continues rather than ending.

Our tool reflows those back into flowing paragraphs by default, using the block structure the engine detects. There is a keep line breaks toggle for when the original layout is the point:

• Reflow for prose — essays, articles, book pages.
• Keep line breaks for receipts, tables, code, addresses, poetry, and anything where the line endings carry meaning.

If your output looks mangled, try the other setting before you conclude the recognition failed.

Where OCR will genuinely let you down

Being honest about the limits saves more time than any tip above.

Handwriting. Standard OCR is trained on printed type. Neat block capitals sometimes come through; ordinary cursive essentially never does. This is a different technology (handwriting recognition), not a matter of trying harder.

Maths and formulae. Superscripts, subscripts, integrals, and Greek letters come back scrambled or silently wrong — and silently wrong is the dangerous part. Equations need specialist tools.

Multi-column layouts. Newspapers, journal PDFs, and two-column conference papers can be read across the columns rather than down them, interleaving two sentences into nonsense. Crop one column at a time.

Tables. You will get the text; you will not reliably get the structure. Cell boundaries, merged cells, and column alignment are usually lost.

Non-English text. Our tool recognises English. Accented Latin text may partly survive; Cyrillic, Greek, Arabic, and CJK will not.

Stylised or historical type. Decorative fonts, heavy ligatures, and old printing — the eighteenth-century long s that looks like an f is a classic — trip it up badly.

Low contrast. Grey text on a grey background, watermarks behind the words, or highlighter over the line all reduce accuracy.

Always check the numbers

This is the tip worth taking away.

OCR mistakes cluster in predictable pairs: 0 and O, 1 and l and I, 5 and S, 8 and B, and rn read as m. In prose you will spot these instantly, because the word looks wrong. In numbers you will not — 1985 read as l985, or a page range of 211–219 read as 2ll–2l9, looks perfectly plausible at a glance.

So if you are transcribing anything where digits matter — dates, statistics, page numbers, DOIs, ISBNs, prices — read those back against the original character by character. Everything else you can proofread at normal speed.

If you are quoting it

Two rules for academic work:

Verify the quotation against the source before you use it. An OCR error inside quotation marks is your error, not the tool's, and misquoting is a serious problem in a way that a typo elsewhere is not.

Get the page number from the original, not from the OCR text. Scans lose pagination constantly, and a quotation attributed to the wrong page is difficult for a reader to check and easy for a marker to notice.

If the source is a published paper, our Citation Generator will build the reference from its DOI, which saves retyping author names and dates that OCR frequently mangles.

In short

Feed it a straight, well-lit, high-resolution crop of printed English text and OCR is genuinely excellent. Feed it a dim photo of a handwritten page at an angle and no tool will save you.

When the output looks wrong, fix the image first — that solves it far more often than switching tools does.