How OCR works: from pixels to text

OCR — optical character recognition — is the trick of turning a picture of writing into text a computer can use. Modern OCR is a neural pipeline, and a small one runs right inside your browser when ToPDF suggests a file name.

The pipeline

1. Clean up: normalise lighting and contrast so strokes stand out (the enhancement you can see helps the OCR you can't).
2. Find text regions and lines: the page is segmented — where are words, how do they group into lines.
3. Read sequences: a neural network slides along each line and emits character probabilities, considering context (an ambiguous glyph between letters is probably a letter).
4. Post-process: assemble words, apply domain rules — for passports, validate check digits; for cards, look for the name-above-birthdate pattern.

Why OCR still misreads

• Fancy fonts and handwriting stray far from training data.
• Low contrast/noise: the classic 0/O, 1/l/I, 5/S confusions bloom when strokes are faint.
• Layout chaos: stamps over text, tables, mixed scripts.

This fallibility is why ToPDF uses OCR only where a mistake is cheap and visible — a suggested file name you can overtype — and why documents themselves are kept as faithful images. (Image vs searchable PDFs.)

On-device OCR, specifically

Cloud OCR ships your text to a server; on-device OCR downloads a compact model once and reads locally. ToPDF bundles a general text model plus a specialist for passport machine-readable zones — both cached in your browser, both offline-capable, neither ever transmitting a character of your documents.

Frequently asked questions

Which languages does the name suggestion handle?

The reader targets Latin-script text (names on IDs and MRZs are Latin by international standard, even on multilingual cards). Non-Latin body text is simply left unread rather than mangled.

Can I get full-text OCR of my scans?

ToPDF deliberately stops at name suggestions. For searchable archives, run a local OCR tool over the PDFs — and keep privacy in mind if a tool is cloud-based.