Incorporating A Rich Linguistic Model into Whole-Book Recognition

Whole-book recognition, a technique that improves recognition of book-images using fully automatic mutual-entropybased model adaptation, has achieved character error rate as low as 1.9% on 50 pages of real book images in our previous publications. However, the linguistic model for word recognition was simple, assuming a uniform distribution on the words in the dictionary, so that the algorithm is unaware of prior word-occurrence distribution. As a result, the statistics of the output transcript differs largely from that of a real distribution. In this paper, we propose a post-processing technique that improves the existing whole-book recognition results by applying the constraints of a rich linguistic model a prior word-occurrence distribution. This technique further drives the character error rate down from 1.9% to 0.97%. We also show that the wholebook recognition algorithm combined with this post-processing technique shows faster improvements in which word error rates fall monotonically with passage length.