Using a Hidden-Markov Model in Semi- Automatic Indexing of Historical Handwritten Records

Indexing of historical records is a process that uses human effort to read text images and convert them into a machine readable format that facilitates search. The Church of Jesus Christ of Latter-day Saints has been using volunteers to index millions of microfilm images of genealogy records collected throughout the world. This indexing process is time-consuming. We adapt a technique for holistic handwritten word recognition originally published by Victor Lavrenko et al., and use it to semi-automatically index US census record images to improve the efficiency of the manual indexing process used in the LDS Church's "Internet Indexing" project. The data used for this project differs from Lavrenko's paper in that we recognized words in three columns of a structured census image instead of unconstrained handwritten words. The approach resulted in 90% accuracy for the chosen columns using as little as one page of manually indexed training data.