Image texture analysis with fast similarity search for content based retrieval and navigation

Kuan, Joseph (1998) Image texture analysis with fast similarity search for content based retrieval and navigation. University of Southampton, Doctoral Thesis.

Abstract

One of the main challenges of multimedia and hypermedia research is the effective use of the media content for retrieval and navigation in multimedia environments. This thesis is concerned with the use of texture as one of the keys for content based retrieval (CBR) and content based navigation (CBN). Other authors have proposed texture analysis procedures and an initial aim was to identify a versatile texture representation which is effective over a very wide range of textures and which could be used efficiently in the context of CBR and CBN. In order to index the multidimensional feature vectors representing texture efficiently, this thesis has also focused on issues of multidimensional indexing for fast similarity search.

This thesis proposes a novel texture representation method which uses the edge and plain region information from texture patterns. The information is used to evaluate contrast across edges, the mean greylevel of plain regions and the conditional probability matrix of edge directions and plain regions as features. A weighted Euclidean measurement for this method is proposed which gives better matching than the standard Euclidean measure. The new representation is compared with a range of previous texture representation schemes using a wide range of texture patterns and its classification properties and speed performance are shown to be an improvement on the other schemes. Since texture is typically represented by a multidimensional feature vector, this thesis investigates multidimensional indexing strategies and knn retrieval methods and proposes new and more efficient approaches in the context of multimedia information handling.

Two different multidimensional indexing approaches are explored in this thesis; the R*-tree and the Hilbert R-tree. Data object retrieval and range search performance are compared in various aspects, including the number of dimensions, nature of databases and database size. k nearest neighbours (knn) similarity search is significant for image based CBR and CBN in multimedia systems, and a new algorithm for knn search is proposed which is an improvement over previous approaches for image based CBR and CBN applications.

The novel texture representation technique, fast indexing R-tree and the enhanced R-tree similarity search technique are integrated into an open hypermedia system which offers content based retrieval and navigation for multimedia data. The thesis concludes with examples of the use of the system for texture based image retrieval and also texture based navigation from images to other media types.

This record has no associated files available for download.

Contributors

Download statistics