<title>Retrieval by content in symbolic-image databases</title>

Abstract

A class of images termed symbolic images is defined. These are images in which the set of objects that may appear in the image is known a priori. Furthermore, the geometric shapes of these objects are relatively primitive, and the information that they convey is mainly symbolic. The main components that are necessary for building a database system that manages and supports retrieval by content of such symbolic images are described. A system that utilizes the symbolic nature of such images to drive symbol recognition for input into an image database is presented. Two approaches for integrating the results of this input system into a spatial relational database (termed the classification approach and the abstraction approach) are suggested. Methods for storing and retrieving images using these two approaches are described. In particular, schema definitions and indices that support spatial, contextual and hybrid queries for each approach are presented. Image retrieval is demonstrated by several example queries specified textually and pictorially. Strategies to answer these queries efficiently, along with detailed execution plans are given. Image similarity is also defined. Algorithms for retrieving images based on a pictorial specification using image similarity are described. Results of an experimental study that compares these approaches are presented. The two approaches are compared qualitatively and quantitatively.