We present a framework to consider the various costs and trade-offs of different methods for working with a large corpus of images containing unstructured text. We illustrate this framework and present performance statistics via a case study identifying racial restrictions in 5.1 million images of Philadelphia property deeds from 1910–1965. We find 8,395 deed images containing a racial restriction, which we estimate to cover more than 6% of all parcels in Philadelphia.

View the Full Working Paper