Spatial and Grayscale Metadata for Similarity Searches of Image Databases

Abstract

This paper presents a content-based image retrieval process wherein the user identifies a feature of interest using a region quadtree decomposition of the image, spatial statistics and histograms of the grayscale values of the feature definition are calculated, and the result is compared to a database of these same calculations that have been performed on similar images. The sum of squared differences between the indices calculated for the quads that form the feature of interest and corresponding quads in the database yields a ranked list of matching image tiles. In an analysis of Landsat 7 imagery of North Georgia and an IKONOS panchromatic image of Kalamazoo, Michigan, we found that the retrieval success rate depends on the spatial and spectral characteristics of the feature of interest and the configuration of quads used to define the feature.