Image interpretation by combining ontologies and Bayesian networks

Abstract

A drawback of current computer vision techniques is that, in contrast to human perception that makes use of logic-based rules, they fail to benefit from knowledge that is provided explicitly. In this work we propose a framework that performs knowledge-assisted analysis of visual content using ontologies to model domain knowledge and conditional probabilities to model the application context. A bayesian network (BN) is used for integrating statistical and explicit knowledge and perform hypothesis testing using evidence-driven probabilistic inference. Our results show significant improvements compared to a baseline approach that does not make any use of context or domain knowledge. © 2012 Springer-Verlag.