multiple objects and feature queries on large image databases. Contents such as color, texture, and shape may be Efficient indexing models play a key role in the contentbased retrieval of image and video databases. In this paper, used for queries, and techniques such as histograms, FFT, four models of image data representations are examined for wavelets have been used to extract relevant parameters automatic indexing from pixel, nearest neighborhood, block and coefficients for indexing [4-6, 8, 11, 12, 14, 22, 30, 34].
to full image. For each model, their representing capabilities,
An index provides a means of identifying an image withinvariant properties (translation, rotation, reflection, and conout involving the full image. In conventional databases, nection) and complexities are assessed. Representing a phase indexing a table or a class consists of either a single field space as a generalized histogram, it is essential for an indexing or combination of several fields directly extracted from the model to have suitable invariants in organizing its phase space. structured information. In terms of size, an index should Considering class histograms, a list of invariant criteria for represent a small fraction of a record, whether the underlyautomatic indexing can be identified. It is interesting that the ing record is highly structured, text-based, or pictorial, and nearest neighborhood model under classification offers the best capability with respect to these criteria. Using the nearest neigh-the smaller the index, the more efficient would be the borhood model, a graph index scheme for 2D-thinned is presearch. An index may be defined as [1] ''a systematic guide sented. This is a typical example how can we combine local to items contained in, or concepts derived from, a collecconnectivity invariants to be global characteristics using a class tion. These items or derived concepts are represented by histogram. From a thinned image, a classification based on entries arranged in a known or stated searchable order, the nearest neighborhood is used. The classification organizes such as alphabetical, chronological, or numerical.'' With possible local patterns into classes with translation, rotation, feature indexing, the aim therefore is to use some informareflection, and connectivity invariant properties. Applying the tion derived from the image, much smaller than the image classification to the full image, the image is partitioned into a itself, to summarize certain features and characteristics of set of feature images. Each feature image contains specific the image for identification purposes. Although manual feature points corresponding to one pattern class. Counting the extraction of image features may be unavoidable in some number of feature points on each feature image as a number, a class histogram can be established. Selecting specific measures situations [23]
, it is always desirable to have a significant of the histogram, three graph indexes can be formulated. Evaluautomatic extraction component that complements the ations of these indexes have been undertaken using capital manual approach. Thus automatic feature extraction and letters and some sample images. These samples illustrate that indexing even at a low-level is always an essential element the present scheme offers a good degree of content discriminain an image database.
tion, and appears to provide a promising framework for the An image is a discrete geometric framework with fixed construction of robust content indexes. Β© 1997 Academic Press boundaries on a given lattice. Different contents of images could be expressed by various shapes and forms regarding their specific geometry [16,18, 20, 32]. An automatic in-
1. Introduction
dexing scheme should be an algorithmic procedure, which establishes a set of quantitative or symbolic measures that Content based image and video databases have seen a rapid progress in recent years, and large-scale content-directly correspond to certain aspects of content-based inbased systems have been built, i.e., QBIC, CANDID, formation for shapes and forms. Different image retrieval TRADEMARK [15,21,25, 28], which support complex systems may apply various indexing schemes to support their content-based image retrieval capabilities. Many researchers are focused on the direction of proposing new