Abstract
This paper describes an image understanding system which reads printed chemical structures via an image scanner, recognizes them as chemical structures and gives the result in the form of connection tables. The system accepts chemical structure diagrams including characters and stereochemical symbols depicted in ordinary textbooks and articles. The recognition and understanding process of image data comprises the following three major steps: at first, the basic recognition of building blocks of input data as figure primitives, then the graph generation from the result of the previous process, and the interpretation and understanding of the graph as a chemical graph (a chemical structure). In other words, the whole process consists of hierarchically distributed functions, where the bottom layer is the first one, and the top the last. The process normally progresses from the bottom to the top. However, the system works cooperatively between the layers when some layer suspends judgements or detects the possibility of misjudgements. For instance, an assumed character area could be ambiguous when the character is overlapped with other figure elements; crossing lines may be interpreted as a cross of two single lines, or four lines connected to a single node. These undetermined situations can be resolved with cooperative processing of two or more layer functions using domain knowledge concerning chemical structures. Experiments have proved that the system is powerful enough to give sufficient precision in terms of center-line extraction of complex chemical structures which a commercial product cannot recognize precisely.
