Fast Retrieval by Spatial Structure in Image Databases

Abstract

The state-of-the-art approach for speeding-up the time responses in databases is using Spatial Access Methods (SAMs) like e.g. R-trees. However, these methods do not treat image content directly (e.g. objects are approximated by their minimum bounding rectangles), nor can they handle images with multiple regions. The proposed approach extends the existing framework of indexing using SAMs to treat image content in conjunction with two well-known image-matching methods, namely the editing distance on Attributed Relational Graphs (ARGs) and the Hungarian method for graph matching. It provides index support for the two most common types of similarity queries, referred to as range and nearest-neighbor queries and has many desirable properties. For instance, it handles even complex queries specifying multiple objects (such as queries by image example), it returns exactly the same answers with the sequential scan methods (without indexing) and works with any SAM (e.g. R-tress) and with any image distance function provided that it satisfies the so-called Lower Bounding Principle .