Image Decomposition and Representation in Large Image Database Systems

To an increasing extent, applications demand the capability of retrieval based on image content. As a result, large image database systems need to be built to support effective and efficient accesses to image data on the basis of content. In this process, significant features must first be extracted from image data in their pixel format. These features must then be classified and indexed to assist efficient retrieval of image content. However, the issues central to automatic extraction and indexing of image content remain largely an open problem. Tools are not currently available with which to accurately specify image content for image database uses. In this paper, we investigate effective block-oriented image decomposition structures to be used as the representation of images in image database systems. Three types of block-oriented image decomposition structures, namely, quad-, quin-, and nona-trees, are compared. In analyzing and comparing these structures, wavelet transforms are used to extract image content features. Our experimental analysis illustrates that nona-tree decomposition is the most effective of the three decomposition structures available to facilitate effective content-based image retrieval. Using nona-tree structure to represent image content in an image database, various types of content-based queries and efficient image retrieval can be supported through novel indexing and searching approaches. We demonstrate that the nona-tree structure provides a highly effective approach to supporting automatic organization of images in large image database systems.