Reading the page as it was printed

A recognition service of our own is now one of the services an OCR job can be set to. Nothing changes on a workflow you already run until you change it there.

Printed and handwritten material go through the same job, with no separate setting to choose between them, and left-to-right and right-to-left material sit in one collection and run through one workflow. Older typefaces, blackletter among them, are read as their own kind of printing rather than forced through a modern-print reading. It is trained on what archives hold — newspapers, books, catalogue cards, forms and manuscripts — rather than on clean office documents, and it chooses how to read each page.

Text comes back with the position of every word on the page. That single property is what makes the rest work: a search inside an item is highlighted where it occurs, a headline or a paragraph can be selected and copied off the image, and a phrase inside a long item is a place you can land on.

Running recognition again over a page that already has a layout keeps the layout and replaces only the words. Zones are not redrawn, articles are not regrouped, picture regions are left alone — so a catalogued, segmented collection can have its text improved in place.

Where a collection’s script, typeface or hand is not covered yet, recognition can be taught it, and the coverage added to an installation already running. It can run offline, inside infrastructure you control; nothing need leave the building.