Optical character recognition

Why it matters more than it sounds#

Scanning a filing cabinet produces image files. Without recognition, those files are findable only by whatever someone typed into the filename, which in practice means they are not findable at all. OCR is the step that converts a digitisation project from a storage exercise into a retrieval capability, and skipping it is the most common reason paperless programmes deliver nothing.

What it reads reliably, and what it does not#

Printed text on a clean scan at 300 dpi recognises very well. Recognition degrades with poor contrast, skew, unusual fonts, tables with ruled lines, and photographs taken at an angle in bad light — which describes most documents captured on a phone.

Handwriting does not recognise reliably, and any vendor implying otherwise is overstating. Handwritten annotations on a delivery note or a clinical record generally stay unsearchable, so the practical approach is to index those documents by structured metadata captured at the point of scanning, rather than relying on full-text search.

Searchable PDF versus extracted text#

Two outputs are commonly confused. A searchable PDF embeds a text layer behind the page image, so the file itself can be searched in a reader. Extracted text is written into a search index, which is what makes a document findable from outside the file. A document management system needs the second; the first is a convenience.

Practical implications#

Run recognition on arrival rather than on demand, so a document is searchable by the time the person who uploaded it has finished filing. Expect to re-run it across an existing archive as a background task after a migration, and expect a proportion of that archive — usually the oldest and worst-scanned part — never to become reliably searchable.