Structure recognition methods for various types of documents

In this paper, we describe experimental methods of recognizing the document structures of various types of documents in the framework of document understanding. Namely, we interpret document structures with individually characterized document knowledge. The document understanding process is divided into three procedures: the first is the recognition of document structures from a two-dimensional point of view; the second is the recognition of item relationships from a one-dimensional point of view; and the third is the recognition of characters from a zero-dimensional point of view. The procedure for recognizing structures plays the most important role in document understanding. This procedure extracts and classifies the logical item blocks from paper-based documents distinctly. We discuss the structure recognition methods for three classes of documents: 1) table-form documents, filled-in forms, cataloging lists, etc. — each item block is surrounded by horizontal and vertical line segments; 2) library cataloging cards, name cards, letters, etc. — each item block is separated by spaces; 3) newspapers, pamphlets, etc. — each item block is constructed hierarchically and by combining under roughly specified layouts. The structure recognition procedure is characterized by individual recognition methods: in class 1 documents, binary trees indicating the connective relationships among neighboring item blocks, which are surrounded by line segments; in class 2 documents, binary trees defining the spatial and geometric relationships among neighboring item blocks, which are separated by spaces; and in class 3 documents, composition rules specifying the constructive relationships among neighboring item blocks, which are represented by adjacent relationship graphs. The methods are effective under the knowledge-based frame-work and are integrated complementarily from the top-down (model-driven) and bottom-up (data-driven) approaches. Of course, the integration means vary according to document classes.