OCR (optical character recognition)

Definition

Reconstructing machine-readable text from an image of a page — the operation that gives a scanned document the text layer it was never born with.

OCR is a recognition problem, not a parsing problem, which is why it is probabilistic in a way the rest of the document pipeline is not. Accuracy depends on scan resolution, contrast, skew, language, typeface and layout complexity. Clean 300 DPI printed text is close to solved. A faxed carbon copy of a handwritten form is not.

That variability has a practical consequence people underestimate: OCR output is evidence of what a page probably says, not a faithful transcript. Search over OCR'd archives should be treated as a recall aid rather than a guarantee, and anything legally consequential should be verified against the rendered page.

OCR also costs meaningfully more compute than rendering does, which is why it is usually a batch operation over an archive rather than something you run when a user opens a file.

In Doconut

OCR ships only through the paid AdvancedSearch plugin, and it is not generally available today — the OCR engine is being re-based. Do not plan around Doconut OCR, and do not describe it as a free or built-in capability. Documents that arrive as scans render correctly; they simply are not searchable.

See how it works in practice

Definitions only carry you so far. A temporary licence runs on your own machine, against your own documents.