Model Based Understanding of Document Images

Document image understanding is a task to generate the structured description about contents of a document. In this paper, we propose a new method of document image understanding which employs the domain specific knowledge base called document model. Document model is structural representation of constraints on the layout structure as well as the logical structure of a target document. Since the variation of the structure can be described in document model, intermediate results of understanding generally include multiple candidates. In order to generate plausible description from such candidates, we introduce the strategy of hypothesis generation and testing. From the experiments for 100 visiting cards, we demonstrate the effectiveness of our method.