DLI-2 Word Spotting: Indexing Handwritten Manuscripts

University Of Massachusetts Amherst, Amherst MA

Investigators

Abstract

This project will research and develop innovative techniques for indexing handwritten historical manuscripts. Automatic indexing of historical archives to create indexes similar to those at the back of most printed books would potentially make available a whole new set of materials to scholars and students. Conversion of printed materials usually involves Optical Character Recognition (OCR) to convert them to machine-readable form. OCR does not work well on handwritten text. The investigators propose to use a scheme known as Word Spotting in which a document page is segmented into words and lists of words are created. By matching word images against each other multiple instances of the same word are then identified. A user then provides the ASCII equivalent to a representative word image from the lists and links to the original documents are automatically generated. For this approach to succeed, a number of problems need to be solved including new techniques for "cleaning up" a document by removing non-meaningful visual artifacts, extending existing algorithms for word segmentation of handwritten documents, and building new algorithms to find similarity between handwritten word images.