Someone sends a photo of a printed page — a recipe, a contract clause, a page from a textbook — and you need the words as text you can search, edit or paste. Retyping it is the obvious option and the worst one.
OCR does this in a couple of seconds. The reason people give up on it is that they photograph the page badly, get mangled output, and conclude the technology does not work. Almost always the capture is the problem, not the recognition.
What OCR is actually doing
Optical character recognition looks at the shapes in an image and decides which characters they are. It is pattern matching against what letters normally look like, plus a language model that knows which words are plausible.
Two consequences follow from that, and they explain nearly every bad result:
- It needs clean shapes. Blur, shadow, glare and skew distort the letterforms, and a distorted letterform is a guess.
- It needs to know the language. A recogniser expecting English will make nonsense of Arabic, and one expecting Latin script will not attempt Arabic at all. Scripts that join their letters, like Arabic, are harder than scripts that do not, and need a model trained for them.
The four things that wreck accuracy
Fix these and most OCR problems disappear:
Shadow across the page. The single biggest one, and the least obvious — your own body blocking the light. Stand so your shadow falls away from the paper, or turn the page toward a window.
Shooting at an angle. A page photographed from the side has letters that are wider at one end. Shoot straight down, or let the app's perspective correction flatten it — that is what document mode is for.
Low contrast. Grey text on cream paper, or a photo taken in dim light with the phone brightening everything into mush. Document mode fixes this by pushing the background to white and the text to black.
A curled or folded page. Letters bend with the paper. Flatten it under something before you shoot; no software fully undoes a curve.
None of this needs a better phone. A mid-range camera on a flat, evenly-lit page beats a flagship photographing a curled page in a shadow.
Getting the text out, step by step
- Capture in document mode, not the camera app. You want edge detection, perspective correction and contrast handling before recognition runs.
- Set the language before you scan if the app asks. This matters most for Arabic and other non-Latin scripts, where the wrong model produces gibberish rather than near-misses.
- Run OCR and read the output, particularly numbers. Digits are where recognisers slip most — a misread digit in an invoice or an ID number is worse than a misread word, because nothing looks wrong.
- Copy or export. Straight to the clipboard for a short passage; to a searchable PDF or a text file for anything longer.
For a multi-page document, scan the whole thing into one file before running recognition rather than doing it page by page — the ordering pitfalls are covered in scanning multiple pages into one PDF.
Arabic, and why it is the awkward case
Arabic OCR is noticeably harder than English. Letters change shape depending on their position in a word, they join up, and diacritics carry meaning. A general-purpose recogniser trained mostly on Latin script handles this badly, which is why so many apps that claim to support Arabic produce something unusable.
If Arabic is your main use, test it before you rely on it: scan one page you can verify and read the output properly. DocFlow Scanner runs Arabic and English recognition on the device, and the wider picture is in free Arabic OCR. If you also need the text in another language afterwards, translating text from an image covers that path.
Where the page goes matters
Most free image-to-text websites work by uploading your image to their server. For a recipe that is fine. For a payslip, a medical letter, a contract or an ID card, you are handing a document to a service whose retention policy you have not read — the reasoning is in are online PDF converters safe.
On-device OCR sidesteps it, and you can verify the claim instead of trusting it: turn on airplane mode and scan anyway. If the text still appears, nothing left the phone.
Frequently asked questions
Can I extract text from a photo I already took? Yes — most scanner apps let you import from the gallery. Accuracy will be whatever the original capture allows, so a shadowed or angled photo still produces poor results. Re-shooting is often faster than fixing the output.
Does OCR work on handwriting? Far less reliably than on print. Neat, separated handwriting sometimes works; ordinary cursive usually does not. Assume printed text unless an app specifically advertises handwriting.
Why are the numbers wrong when the words are right? Because the language model helps with words and cannot help with digits — there is no "plausible word" check for a random number. Always proofread figures, especially in invoices and ID numbers.
Do I need internet for OCR? Not if the app processes on the device. If it stops working in airplane mode, it was sending your page to a server.
What is the best format to keep the result in? A searchable PDF keeps the original page image with the text layered underneath, which is usually what you want for documents. Plain text is better when you only need the words.
The bottom line
The capture decides the result: flat page, even light, no shadow, shot straight down, in document mode with the right language set. Do that and recognition is close to effortless; skip it and no app will save you. Proofread the numbers, and if the page is anything sensitive, use a scanner that runs on the device — which is what DocFlow Scanner does for both Arabic and English. Still choosing an app? The free Android scanner roundup lists the checks worth running first.
