How to Extract Text From an Image or Screenshot (Free OCR Guide)

Someone sends a photograph of a page. A form arrives as a screenshot. A whiteboard full of notes exists only as a picture on your phone. The information is right there and completely unusable, because you cannot select a single word of it.

Optical character recognition solves this, and it works far better than its reputation suggests. It also fails in specific, predictable ways that are worth knowing before you trust the output.

What OCR actually does

OCR does not read. It finds regions of an image that look like lines of text, isolates individual character shapes, and matches each shape against what it has learned characters look like. Every character that comes out is a best guess with a confidence score attached.

That framing explains everything else. Anything that makes a character shape clearer raises accuracy, and anything that distorts the shape lowers it. The engine has no idea what your document means, so it cannot catch a mistake by noticing that a sentence stopped making sense.

First, check that OCR is worth it

For a short passage, retyping is often faster and is guaranteed correct. OCR earns its place on volume: a full page, a long table, a stack of receipts, or anything you would otherwise copy character by character while checking every line.

The other case where it clearly wins is text you cannot type accurately anyway, such as a serial number, a long reference code, or a technical string where a single wrong character makes the whole thing useless.

Screenshots are easy, photographs are not

These two jobs look the same and are not remotely equal.

A screenshot is text that was rendered by a computer: perfectly sharp edges, even spacing, no shadow, no distortion, high contrast. Accuracy is usually near perfect, and if it is not, the image was probably scaled down before you got it.

A photograph of paper arrives with perspective distortion because the camera was not perfectly square to the page, uneven lighting, the curve of a bound page, shadow from your own hand, and often motion blur. Every one of those deforms the character shapes the engine is trying to match.

What actually decides accuracy

• Resolution. The single biggest factor. Text needs enough pixels per character to have a recognisable shape. Around 300 DPI for scanned paper is the usual guidance, and a photo cropped tightly to the text beats a wide shot at the same megapixels.
• Contrast. Dark text on a light background is what engines are trained on. Grey text on grey, low-contrast screenshots, or a photo taken in dim light all cost you accuracy.
• Straightness. A page rotated a few degrees is handled well. A page photographed from an angle, so the far edge is narrower than the near one, is much harder.
• The typeface. Ordinary print fonts are easy. Decorative, condensed, script, and very small fonts are considerably harder, and handwriting is a different problem altogether.
• Layout. A single column of text is straightforward. Columns, sidebars, and tables force the engine to decide reading order before it can decide anything else.

How to extract the text, step by step

1. Open the image to text converter and add your picture.
2. Let the engine load. The recognition model is fetched on first use and then cached, so the first run is slower than the ones after it.
3. Wait for recognition, which happens in your browser rather than on a server.
4. Download as DOCX if you want to keep editing, or as plain text if you are pasting it somewhere else.
5. Proofread before you use it. Read the numbers first, then the names, then everything else.

How to rescue a bad result

If the output is garbled, do not try a different tool first. Fix the picture, because that is almost always what is wrong.

• Crop to the text only. Remove the desk, the fingers, and the margins. It focuses the engine and raises effective resolution.
• Straighten the page so the lines run horizontally.
• Retake it rather than enhance it. Phone flat above the page, room light in front of you, no shadow from your own body, and as close as you can get while keeping the whole text block in frame.
• Split multi-column pages into one image per column, then join the text afterwards. This single trick fixes most interleaved-nonsense results.
• Do one page at a time. Two pages in one photo halves the resolution of each.

What OCR still gets wrong

Even a clean run has characteristic failure modes, and they are the ones that matter most.

Similar characters. The digit 1, lowercase l, and uppercase I are near-identical in many fonts, as are 0 and O, 5 and S, and 8 and B. These substitutions are silent and read perfectly naturally.

Numbers generally. A misread letter in a sentence is obvious. A misread digit in an invoice total, an account number, or a dosage is invisible and consequential. Always check figures against the original by eye.

Structure. Tables usually come out as a stream of values with the grid gone, and columns can be interleaved. The words are right; the arrangement is not.

Handwriting. Print-style handwriting sometimes works. Cursive generally does not, and no amount of image improvement changes that.

Scripts it was not trained for. Recognition models learn the shapes of a particular writing system, so a model built around Latin characters will not read Devanagari, Arabic, or Cyrillic well, and a page mixing two scripts is harder than either alone. If your document is not in the script the engine expects, treat the output as a draft rather than a copy.

Worth knowing: OCR rarely announces failure. It returns its best guess at every position, so a badly misread page still arrives looking like ordinary text. Silence is not confirmation that it worked.

A proofreading order that catches the most

Read the output against the original in a fixed sequence rather than straight through. Numbers first, because they are both the most likely to be silently wrong and the most costly. Then proper nouns, which have no surrounding context to correct them. Then anything in unusual formatting, such as headings and captions, where the engine had least to work with. Ordinary sentences come last, because you will notice an error there while reading anyway.

If it is a PDF, check before you OCR

Many PDFs already contain real text and need no recognition at all. Try selecting a line: if the words highlight, use a PDF to Word converter instead and you get the characters exactly as typed, with no recognition errors possible. This guide to converting PDF to Word covers the difference in detail. Only reach for OCR when the page is genuinely an image.

Afterwards

Once you have the text, a word counter is a quick sanity check that nothing was dropped, and text to PDF turns the cleaned-up result back into a shareable document.

One last point about the pictures people run through OCR: they are overwhelmingly identity documents, invoices, prescriptions, and contracts. A tool that recognises text inside your browser never uploads any of that, which is the right default for material you would not email to a stranger. Fix the image, run it once, and read the numbers twice.