Image retrieval based on colour and nonlinear texture invariants

Abstract

Among other techniques especially methods working fully automatically are of interest for image retrieval from large databases. Colour histograms proved to be successful in automatic image retrieval, however, their drawback is that all structural information is lost. Therefore we extend the colour histogram approach by features that take into account the relations within a local pixel neighbourhood. By integrating nonlinear functions over the group of Euclidean motion we extract features that are invariant with respect to translation and rotation. In contrast to approaches using linear filtering (e.g. wavelets) or corresponding power spectra (to become invariant) these nonlinear invariants have the potential to be unique with respect to the equivalence class of Euclidean motion. So in invariant feature histograms we combine the advantage of an invariant description (e.g. we only need one histogram for a whole class of transformed images in the database) with the properties of histogram approaches, providing the possibility to find images also by partial views or vice versa or to detect objects also under occlusion.