Matching structural descriptions of handwritten characters using heuristic graph search

Structural representations are widely used in object recognition. They model the parts that form an object and the relationships between those parts. We discuss techniques for deriving and comparing structural representations of handwritten characters. Character recognition can be modeled as two main tasks; to derive an abstract description of a character from the raw data describing the writing process, and to find an association between a character's abstract description and the identities of characters known to the system. Representations are selective in the information they contain and the emphasis they place on aspects of the information. The nature of the abstract representation used is crucial to the recognition process. It constrains the information available to the processes responsible for establishing a character's identity. There are a number of desirable properties of a representation It should be compact, rich, extensible, and robust against deformations and noise. In addition it should be stable, such that small changes in raw data cause only proportionate changes in the abstract representation. Structural descriptions go some way to meeting these criteria. The article relates to online cursive handwritten characters drawn from an alphanumeric character set. We look at the process leading to the formation of a structural description and its implementation using a graph representation. We then describe a process for comparing graph representations framed as an optimisation task using a graph distance metric.