Diplomarbeit Image Retrieval in the Compressed Domain Using JPEG 2000

This thesis shows the development process of JPEG2000 content based image retrieval (CBIR) methods. First, foundations of CBIR as well as JPEG2000 are explained and existing solutions are examined. On this basis, own approaches are developed. Wavelet coefficients that are extracted from JPEG2000 are used to create different feature sets. To facilitate the development process, the Java program ”JP2FeatureFinder” is developed. The distribution of wavelet subband coefficients is modelled as histogram, Generalized Gaussian Density (GGD) and Gaussian Mixture Model (GMM). In order to decide which images are similar to each other, these distributions are compared using Histogram Intersection (HI) and the Kullback-Leibler Divergence (KLD). Color and texture are treated separately. When merging the individual results, they can be weighted independently. A plug-in for the GNU Image Finding Tool (GIFT) is implemented. Using the GIFT framework, the retrieval methods developed are tested with the TSR2500 image collection consisting of images from the Swiss TV channel Télévision Suisse Romande and ground truth data from the Viper group of the University of Geneva.