What Is OCR and When Do You Need It?
OCR (Optical Character Recognition) reads the letters inside an image and turns them into real, editable text you can copy and paste. It is the technology behind copying a phone number from a screenshot, pulling a quote out of a photographed page, or digitising a scanned receipt. Any time text is locked inside a picture instead of being selectable, OCR is what frees it.
Common situations where OCR saves time:
- Screenshots. Copying an error message, a code or a paragraph out of a picture someone sent you. This is the easiest case there is, because a screenshot is already crisp digital text.
- Receipts and invoices. Pulling totals, dates and reference numbers into a spreadsheet without retyping them. Check the digits afterwards; numbers are where OCR errors hurt most.
- Business cards. Turning a stack of cards into contacts. Names and unusual spellings need a proof-read, but addresses and emails usually come through cleanly.
- Scanned documents. Making an old contract or letter searchable and editable rather than a flat picture of a page.
- Whiteboards and slides. Capturing notes from a meeting photo or a conference deck. Angle is the enemy here, so shoot square on rather than from your seat.
- Infographics and charts. Getting the labels and figures out of an image that has no underlying data file.
How to Extract Text From an Image Online
You do not need to install anything. Our free Image to Text tool runs OCR directly in your browser:
- Open the Image to Text tool and upload your image or screenshot (JPG, PNG, or WEBP).
- Pick the language of the text. English is the default, and over 100 languages are supported.
- Let the tool scan the image. It uses the open-source Tesseract engine running locally on your device.
- Copy the extracted text, or download it as a TXT file.
How to Get the Most Accurate Results
The usual advice is to improve contrast, sharpen the image and clean it up. We tested that advice rather than repeating it, by taking one 189-character passage of invoice text, damaging it twelve different ways, and scoring every result against the original. The full table of results is on the tool page, and it reorders the priorities substantially.
Resolution is the thing that matters. The same passage scored 100% as a clean 1200-pixel screenshot and 7.9% as a 264-pixel-wide thumbnail. At that size the strokes that distinguish an e from a c physically are not in the file, and no amount of processing puts them back. If you can retake or rescan at a larger size, do that before anything else. For scanned pages, 300 DPI is the usual working minimum.
Angle matters, and less than you would think until it does not. A mild phone photo with 2.5 degrees of skew still scored 100%. A harder photo combining 6 degrees of skew with vignetting, noise and slight defocus collapsed to 7.9%. Small tilts are handled; a genuinely awkward photograph is not. Straighten it or shoot it again square on.
Contrast matters far less than everyone says. Our faint test was mid-grey ink on light-grey paper, a 12% luminance gap, the sort of faded scan people apologise for sending. It scored 100%. Do not spend time boosting contrast on a legible image; spend it on resolution.
Compression artefacts are close to irrelevant. The same screenshot saved as a quality 18 JPEG, which looks visibly degraded, also scored 100%. A heavily compressed screenshot is not the reason your OCR failed.
Cropping still helps. Not because clutter confuses the engine, but because it lets you keep the resolution where the words are. Our crop tool is the right first step on a photo of a page taken from a distance.
The short version: fix size and angle, ignore contrast and file quality.
Printed, Handwritten and Multi-Language Text
Printed text is what this technology was built for, and on clean printed material it is close to perfect. Our clean screenshot, our faint scan and our full page at 5000 by 7000 pixels all came back with zero character errors.
Handwriting is a different problem. Tesseract is trained on typefaces rather than on handwriting, and the further a script gets from print, the worse it does. A print-style handwriting typeface scored 98.4% in our test, while a formal cursive typeface fell to 90.5%. Both of those are typefaces designed to imitate handwriting, not real handwriting, which varies far more. We have not tested genuine handwritten notes, so we will not put a number on them; treat handwriting as a partial result to be checked rather than a transcription to be trusted.
For other languages, select the matching model before running the scan. The effect is more interesting than a simple accuracy gain: a Spanish passage scored 98.5% with the Spanish model and 98.5% with the English model too. The language pack changed which errors appeared rather than how many. The English model lost accented characters and inverted punctuation; the Spanish model handled those and then mangled an email address instead. Pick the model that matches the content that matters most to you.
Mixed-language documents are the awkward case, since only one model runs at a time. Run the scan twice with different models and take the better parts of each.
Tables, Columns and Layout
OCR extracts words, not structure. That distinction causes most of the disappointment people feel with the results.
On a multi-column page such as a newspaper or an academic paper, the engine may read straight across the columns, interleaving two unrelated sentences into one line. On a table, the rows and columns arrive as a stream of values with the alignment gone, so a price can end up next to the wrong item. Neither is a bug; the output format is plain text, and plain text has no columns.
Two practical workarounds. Crop each column and run it separately, then join the results in order. And for tables, expect to rebuild the structure by hand and use OCR only to avoid retyping the values, checking every number as you go.
What OCR Cannot Do Well
The single most important limitation is not the error rate, it is the way failure presents itself. OCR does not tell you when it has failed. It returns confident, plausible-looking text either way, and a misread digit looks exactly like a correct one. In our tests a 0 came back as a 9 inside an otherwise clean line.
That has a direct consequence for how you use it: always read the output against the image when the content is numbers, codes, dates or anything financial. For prose, a quick skim is enough, because a wrong word usually reads as nonsense. For a reference number, nothing in the output will look wrong.
Beyond that, OCR will not preserve formatting, will not reliably read text over busy photographic backgrounds, and cannot recover text that was never legible.
Privacy: Where the Image Actually Goes
Most online OCR services upload your file to a server, run recognition there and send the text back. That is a reasonable engineering choice and a genuine privacy question, because the documents people run through OCR are exactly the sensitive ones: contracts, payslips, medical letters, bank statements, identity documents.
Our tool loads the Tesseract engine into your browser and runs the recognition on your own machine, so the image is never transmitted. The practical test costs nothing: open your browser network panel before scanning and watch what leaves. Our image privacy page walks through doing exactly that, and it is worth doing on any tool you plan to feed a real document to.
The one honest trade-off is speed. Local recognition depends on your device rather than on a server farm, so a very large page takes a few seconds rather than being instant. Our 140-megapixel test page took roughly six seconds and still scored 99.5%.
Summary
OCR turns pictures of text back into text, and it is genuinely excellent on clean printed material. Give it resolution and a straight angle and it will usually be perfect; give it a thumbnail or a badly angled photograph and it will fail completely rather than gracefully. Check numbers by hand, expect layout to be lost, and keep the file on your own machine when the document is sensitive. Start with the Image to Text tool.
