To a computer, a scanned page is a grid of colored dots. The letter "A" in a photograph of a paragraph isn't a letter — it's a pattern of pixels that happens to look like one to human eyes. OCR, short for optical character recognition, is the technology that closes that gap: it takes an image of text and works out what the text says, converting a picture of words back into words.
If you're here to run the process on a specific file, the companion guide How to Make a Scanned PDF Searchable With OCR covers the workflow step by step. This post is the other half — what the technology actually is, how it works, and why it sometimes doesn't.
The problem OCR solves
Text is born in two ways. A document exported from Word or a website contains real characters — stored as text, searchable, copyable, indexable. A scanned document contains none of that. The scanner records what the page looked like, and the appearance of text is not text.
That distinction has consequences: you can't search a scan, can't copy a quote out of it, can't hand it to a screen reader, and desktop search tools index it by filename only. OCR exists to restore what the scan lost — the function of the text, without touching its appearance.
Step one: finding the text
Recognition starts before any letter is read, with segmentation: the software divides the page into regions, regions into lines, lines into words, and words into candidate characters. On a clean printed page this is almost mechanical. On anything else — two columns, a table, a stamp across a paragraph — the engine has to reason about which marks belong together.
This stage is also why a crooked page hurts so much: the boxes get drawn at the wrong angles, characters bleed across boundaries, and everything downstream inherits the damage.
Step two: recognizing characters
For each candidate, the software compares what it sees against what it knows about letter shapes. Classic systems matched against templates — stored ideals of each letter — and measured features: stroke counts, holes, curves, where lines meet. An "O" is a closed curve with one hole; a "C" is an open one; an "E" is three horizontal strokes hung on a vertical spine.
What makes this hard is that the same letter never looks the same twice. Fonts vary, sizes vary, ink spreads, paper texture interferes. So the matching is probabilistic: each shape yields a set of guesses with confidence levels, not a single answer.
Step three: making sense of it
The guesses get settled by context. Language models — statistics about which characters follow which — weigh the possibilities and pick the most probable reading of the word as a whole. This is why an OCR engine can sometimes read a smudged word you'd struggle with yourself: it isn't only looking at the shape, it's also asking what would make sense there.
The classic confusions are exactly the shapes that coincide: zero and capital O, one and lowercase L, eight and B, five and S. Context resolves these in real words — a dictionary has "boss" but not "8oss" — which is also why the errors that survive concentrate in numbers and names, where there's no dictionary to lean on.
What makes OCR hard
- Handwriting — there is no stable template; every writer's letters differ, and one writer's differ every time.
- Skew — a page fed at an angle corrupts the segmentation stage before recognition even begins.
- Low contrast and blur — faint print, shadows, and low-resolution scans leave the engine choosing between near-identical shapes.
- Unusual fonts — decorative or condensed typefaces sit far from the letterforms engines learn from.
The flip side: clean, high-resolution scans of ordinary printed text are the easy case — and that's what most documents are.
How OCR shows up inside a PDF
In a PDF, OCR's output is usually an invisible text layer: the recognized characters, positioned at the same coordinates as the image they came from, placed behind it. The page still looks exactly like the scan — but searching now highlights the right phrase, selection grabs a sentence, copy produces real text, and a screen reader has something to read.
That placement precision isn't cosmetic. Because the text sits where the image's text sits, search results and selections line up with what you see rather than approximating it.
Where you run into OCR
Besides searchable scans, the same machinery is what lets you deposit a check by photographing it, what data-entry systems use to read invoices and forms, and what archives use to digitize books. Anywhere a photo of text needs to become text, OCR is somewhere in the pipeline — usually invisibly.
The honest limits
OCR is good, not magic. Accuracy is strong on clear, high-resolution scans of printed text, and it degrades on handwriting, skewed pages, low-quality scans, and unusual fonts. An engine that runs entirely on your own device — like the one behind OCR PDF — is genuinely useful on ordinary documents but won't match paid server-side engines on difficult ones. And this release recognizes English only.
For documents where an error matters — legal, financial, medical — treat recognized text as a draft and check the critical numbers against the page. Recognition software confuses things a human never would, and those slips hide exactly where proofreading is hardest.