How OCR works on iPhone

A scanned page starts life as an image: a grid of pixels that happens to look like writing. To a computer there is no difference between a scan of a contract and a photograph of a wall. You cannot search it, you cannot copy a phone number out of it, and you cannot ask a machine what it says.

Optical Character Recognition closes that gap. It examines the image, identifies the shapes that are characters, works out which characters they are, and produces text. In a PDF that text is then stored as an invisible layer sitting exactly on top of where the words appear in the image — which is why you can select a sentence in a scanned document and have the selection line up with the printed words.

What happens between the photo and the text

Modern recognition runs in roughly four stages, all of which complete in well under a second on current iPhone hardware.

• Detection. The system first finds where text is on the page, marking out regions and lines. This is why a page with two columns generally comes out in the right reading order rather than reading straight across the gutter.
• Normalisation. Each detected line is straightened and scaled to a consistent size, so that a line photographed at a slight angle is presented to the recogniser as though it were level.
• Recognition. A neural network converts each line image into a sequence of characters, along with a confidence score for each.
• Language modelling. The raw character sequence is checked against what words plausibly exist in the target language. This is the stage that turns rn back into m and rescues 1nvoice as Invoice.

That last stage is also why telling the app the right language matters. A Turkish document processed with an English language model will lose its diacritics — ğ, ş, ı, ö, ü, ç — because the model does not expect them and quietly substitutes the nearest plausible English letter.

Why on-device recognition matters

Every scanning app faces the same architectural choice: run recognition locally, or upload the page to a server and get text back. The difference has consequences that go well beyond speed.

The privacy dimension is the one worth pausing on. The documents people scan are, almost by definition, the sensitive ones: contracts, payslips, medical results, identity documents, tax filings. Cloud OCR means those pages are transmitted to and processed by a company you have no relationship with, under a retention policy you have probably not read. On-device recognition removes that question entirely — there is nothing to transmit, so there is nothing to retain, leak or subpoena.

PDF Scanner uses Apple's Vision framework, which runs entirely on the device's own neural hardware. Our privacy policy is short for the same reason: we have no server that ever sees your pages.

Getting accurate results

Recognition quality is decided before you press the button, not after. In rough order of impact:

• Sharpness. Blur is fatal. A steady capture matters more than an expensive camera — brace your elbows, or rest the phone edge on a book.
• Even lighting. A shadow across the page produces a band where contrast collapses and characters become ambiguous.
• Resolution on the text. Fill the frame with the page. A document photographed from across the desk may look fine to you but leaves each letter only a handful of pixels tall.
• Contrast. Apply a black-and-white filter before running recognition on text-only pages. Pushing the paper to white and the ink to black is exactly what the recogniser wants.
• Flatness. Curl near a book spine distorts letter shapes progressively along the line, which is one of the harder distortions for a recogniser to undo.

What OCR is still bad at

It is worth being honest about the limits, because knowing them saves time.

Handwriting is hit and miss. Print-style handwriting in dark ink on white paper often works; cursive rarely does; pencil almost never does.

Decorative and condensed fonts — certificates, invitations, some logos — trip up both the recogniser and the language model, because the letterforms are unusual and the words are often proper nouns the model cannot verify.

Text over images, such as a caption printed across a photograph, loses the contrast the detector relies on to separate characters from background.

Tables are recognised as text but not necessarily as structure. You will get the numbers; you may not get them in the right columns.

Very small print, below roughly 6pt at typical capture distances, simply does not carry enough pixels. Move closer and scan the fine print as its own page.

What you can do once a document is searchable

The payoff is not the text itself, it is what the text makes possible. A library of recognised documents can be searched by content: the amount on a receipt, the name of a clinic, a policy number, a date. You stop needing to remember what you called a file, because you can find it by what it says.

That changes the filing problem from a naming discipline into something much more forgiving — which is the subject of organising documents on iPhone.

Frequently asked questions