SUPERVISED TEMPLATE ESTIMATION FOR DOCUMENT IMAGE DECODING

An approach to supervised training of character templates from page im ages and unaligned transcriptions is proposed. The template training p roblem is formulated as one of constrained maximum likelihood paramete r estimation within the document image decoding framework. This leads to a three-phase iterative training algorithm consisting pf transcript ion alignment, aligned template estimation (ATE), and channel estimati on steps. The maximum likelihood ATE problem is shown to be NP-complet e and, thus, an approximate solution approach is developed. An evaluat ion of the training procedure in a document-specific decoding task, us ing the University of Washington UW-II database of scanned technical j ournal articles, is described.