Compressed Domain Image Retrieval Using JPEG2000 and Gaussian Mixture Models

We describe and compare three probabilistic ways to perform Content Based Image Retrieval (CBIR) in compressed domain using images in JPEG2000 format. Our main focus are arbitrary non-uniformly textured color images, as can be found, e.g., in home user image collections. JPEG2000 offers data that can be easily transferred into features for image retrieval. Thus, when converting images to JPEG2000, feature extraction comes at a low cost. For feature creation, wavelet subband data is used. Color and texture features are modelled independently and can be weighted by the user in the retrieval process. For texture features in common databases, we show in which cases modelling wavelet coefficient distributions with Gaussian Mixture Models (GMM) is superior in to approaches with Generalized Gaussian Densities (GGD). Empirical tests with data collected by non-expert users evaluate the usefulness of the ideas presented.

• Type: Conference
• Year: 2005
• Where: VISUAL
• Authors: Alexandra Teynor, Wolfgang Müller, Wolfgang L. J. Kowarschick
• Keywords: Content Based Image Retrieval, Image Retrieval, Jpeg2000, VISUAL 2005