Bayesian learning and reasoning for context exploitation in visual information retrieval

This paper presents a semantic context inference approach on the basis of a multi-feature based visual information retrieval framework. This approach aims at assisting effective retrieval of visual content by exploiting the context information in the digital database. Bayesian networks are used as an inference tool, which can be automatically constructed by learning from the multi-feature similarities and a small amount of training data. The idea is to model potential semantic descriptions of basic semantic concepts in the visual content, the dependencies between them, and the conditional probabilities involved in those dependencies. This information is then used to calculate the probabilities of the effects that those concepts have on each other in order to obtain more precise and meaningful semantic labels for the visual content. However, the proposed method is not restricted to the specific multi-feature based visual information retrieval framework used in this paper. Selected experimental results are presented to show how the proposed context inference approach could improve the retrieval performance.