Kovalev, V.V.KovalevVolmer, S.S.Volmer2022-03-092022-03-091998https://publica.fraunhofer.de/handle/publica/33022910.1109/MULMM.1998.722972Multimedia documents are different from traditional text documents, because they may contain encodings of raw sensorical data. This fact has severe consequences for the efficient indexing and retrieval of information from documents in large unstructured collections (e.g. WWW), because it is very difficult to automatically identify generic meanings from visual or audible objects. A novel method for image retrival from large collections is proposed in this paper. The method is based on color co-occurrence descriptors that utilize compact repressentations of essential information of the visual image content. The set of descriptor elements represents "elementary" color segments, their borders, and their mutual spatial distribution on the image frame. Such representation is flexible enough to describe image scenes ranging from simple combinations of color segments to high frequency color textures equally well. At the retrival stage the comparison between a given query descriptor and the dat abase descriptors is performed by a similarity measure. Image descriptors are robust versus affine transformations and several other image distortions. The consideraton of the descriptors as sets of elements allows the combination of several images or subimages into a single query. Basic properties of the method are demonstated experimentally on an image database containing 20000 images.enCo-occurance descriptorscontent-based image retrievalFeature DescriptionMultimedia Information System006conference paper