Reading constrained text using hierarchical hidden Markov models

Text on documents can generally be classified into one of two categories, unconstrained and constrained, based on the spatial and contextual knowledge that is needed for comprehension. These constraints are found to be critical in text understanding. A model-based approach to reading constrained text is described. The system consists of two major components: a knowledge acquisition module and an image decoding module. Constraints on an application domain are encoded in a hierarchical stochastic model. Subparts of the hierarchy form hidden Markov models, which are used to determine decoding credibility. Verification of a terminal component of the hierarchy is performed by word recognition. Interpretation of a text block is by finding the highest credibility path in the hierarchy. The decoding process is application independent, as demonstrated in the experiments. The system has been tested on reading postal permits and other domains will be used in future tests