Learning qualitative relations in real world scenes

Learning from visual scenes is an innate ability of human beings yet an unresolved and highly complex task for machines and is being addressed in the area of cognitive vision systems. Most of the present developments in cognitive vision systems adopt quantitative or model base approaches. Quantitative approaches involve tedious calculations and yield black box type of learning while in model base approaches the structure of the model has to be known before hand, which is not practical in all situations. Therefore, this research work primarily adopts a qualitative approach in learning from visual scenes and inductive logic programming is used to learn rules of a visual scene from a pool of example scenes. The prototype system developed is capable of learning accurate rules, which in turn used for learning new models and the system is capable of learning beyond initial specifications.