Heuristic similarity measure characterization for content-based image retrieval

Similarity measures are functions which describe the degree of likeness between two objects. We propose a method of using domain-specific expert knowledge to infer a functional similarity measure between objects in that domain. Using a user interface and a collection of exemplar objects, an expert interactively constructs the similarity structure of the domain under consideration. From the expert rankings and dissimilarity assignments, a vector representation of the exemplar objects is found. A neural network is then trained to find a similarity measure, which then can be used for indexing and content based retrieval. Using this approach, a system for retrieval of simple three-dimensional polyhedra is implemented.