OCR Redaction and Text Recognition

Text appearing in video frames, images and scanned documents is recognized and indexed. Each text region is classified by script and routed to the recognition engine that handles it, so a document mixing scripts is read rather than partly missed.

Four OCR engines ship, and the engine is selected per content type, because the engine that suits a video frame is not the one that suits a document page.

Scripts routed

Routing is per region, not per document, so a page mixing scripts is handled region by region.

Objects inside documents, not only text

Document pages are rasterised so the same detection that runs on images runs on the page content. A face, license plate or vehicle embedded as an image inside a PDF is detected and obscured the same way it would be in a standalone image.

This is the gap most document redaction tools have. A text-only tool redacts character runs; it cannot see a photograph on page four.

Tables and spreadsheets

Spreadsheets and tables in documents are redacted by column and by row, so a whole field of sensitive values is removed in one action rather than cell by cell.

Searchable afterwards

Text that appears on screen or on a scanned page is searchable across the library. A result opens at the exact timestamp in a video or the exact page in a document.

Thresholds

OCR confidence thresholds are set separately for video, image and document, because the three have different noise floors.

• Screen recording redaction — on-screen text in dashboards and application windows
• Medical and clinical documents — scanned forms and handwritten notes
• eDiscovery — production sets of mixed scanned and native files
• Legal

• Script routing governs text recognition. PII language coverage is a separate matter and a narrower set. Never read a script list as a PII-detection language list.
• Detection on embedded objects operates on the rendered page, so an object must be visible in the page as rendered.
• OCR quality follows scan quality.

• Text regions are classified by script and routed to the engine for that script, so non-English documents redact.
• Routing is per region, so a document mixing scripts is handled region by region.
• Perso-Arabic coverage spans Arabic, Urdu, Sindhi, Dari and Pashto.
• Four OCR engines ship, selected per content type by deployment configuration.
• Detection runs on the rendered page, so faces, plates and vehicles inside embedded images are covered, which extends beyond the character runs a text-only tool redacts.
• Spreadsheets redact by column or by row.
• OCR text search finds a word that appears on screen rather than spoken, landing on the timestamp or page.