In this review, the subfield of visual interpretation and understanding is first defined and three major issues for using knowledge to increase the functionality and performance of vision systems are introduced. These selected issues concern the role of context, control and learning. In section 2, four approaches to reasoning are distinguished and illustrated with key papers on 1) constraint-based vision, 2) model-based vision, 3) formal logic, and 4) probabilistic frameworks for visual interpretation and control. In section 3, exploitation of these techniques is discussed for automating linguistic descriptions of scenes, enhancing human-computer interaction in multimodal and multimedia systems, in behavioural control for robotics, advanced surveillance systems and biomedical image analysis systems. Finally, promising directions for future research are suggested. These use the deformable models, dynamic learning, and situated approaches to visual understanding discussed in the main sections of the report.
Download compressed postscript file