Shuichi Tsujimoto

Understanding multi-articled documents

A document understanding method based on the tree representation of document structures is proposed. It is shown that documents have an obvious hierarchical structure in their geometry which is represented by a tree. A small number of rules are introduced to transform the geometric structure into the logical structure which represents the semantics. The virtual field separator technique is employed to utilize the information carried by special constituents of documents such as field separators and frames, keeping the number of transformation rules small. Experimental results on a variety of document formats have shown that the proposed method is applicable to most of the documents commonly encountered in daily use, although there is still room for further refinement of the transformation rules.<<ETX>>

Resolving Ambiguity in Segmenting Touching Characters

This paper presents an efficient and powerful character segmentation method which enables touching characters in a document to be read accurately at high speed. The character segmentation phase extracts characters from a text line. Connected components in a text line image may have to be segmented or combined to form recognizable characters. For example, the character ‘i’ is formed by combining two components. Touching characters are segmented to identify each character. Segmenting touching characters is an open problem, whose solution would advance the field. Touching characters have several candidates for their break position, which are then confirmed by recursive segmentation and recognition, and finally by the determination of the linguistic context. There are several possible candidates at each stage. For example, several candidates for the break position of touching characters are nominated. Any segmented area might possibly fit several alternative characters. Therefore, an efficient resolution of ambiguity at each stage is significantly critical and indispensable for practical text reading. The authors’ approach is based on knowledge about character composition (e.g. an ‘m’ is like a combination of an ‘r’ and an ‘n’), as well as knowledge about omni-fonts. Knowledge about character composition compresses the number of recursive segmentation and recognition.

Automatic fixing of ship position by simulation-and-matching

In this paper, we propose an automatic method, named simulation-and-matching, for fixing ship position both from a radar echo image and from a synthetic echo image which is synthesized from three-dimensional topographic data, and we show that simulation-and-matching is a more accurate and robust method than those methods that are based on coastline matching. For bulky three-dimensional data, data compression by spline approximation is also considered.

A common dictionary system shared among multiple media conversion functions

The authors propose a common dictionary shared among various types of media conversion functions. When a new word is registered in the common dictionary, it is immediately available to all functions making use of the dictionary. By editing the common dictionary through a common GUI, users can enter or delete a new word. They have developed an experimental system, which uses a common GUI to enable users to maintain individual dictionaries of natural language processing tasks collectively. A common dictionary maintains consistency across different functions, thus decreasing the need for dictionary maintenance and reducing editing requirements.