In this work we investigate detection of high-level concepts in multimedia content through an integrated approach of low-level visual analysis and exploitation of high-level visual context. Visual context deals with two aspects, namely high-level concepts and low-level features of an image, employing a model of a priori specified semantic relations among the concepts and automatically extracted topological relations among visual features, thus combining both conceptual and topological context. Indicative fragments of proposed "fuzzified" ontologies are shown below:

Fragments of corresponding visual context ontologies.

In our current approach all relations between concepts are represented in such an ontology, each one of them forming a part of the overall contextual knowledge. We focus on a representative example, i.e. the partOf relation P. The partOf relation, P, is a fuzzy partial ordering on the set of concepts. P(x,y)>0 means that y is part of x. For example x could be tennis and y could be a tennis ball. Its representation may take the form presented in the following figure; two concepts are provided and supposing an RDF namespace &dom following sample part of RDF representation is produced:

RDF representation of the partOf relation.

A set of algorithms is also developed, which modify either the confidence values of detected concepts, or the model vectors of low-level features based on which visual detection is performed. Some indicative results from the application of visual context on a set of regions of one sample image containing six possible high-level concepts are presented in the following figure.

Indicative contextualization results.