High-level vision is concerned with constructing a model of the visual
world and making use of this knowledge for later recognition. In this
paper, we develop a general framework for representing uncertain know
ledge in high-level vision. Starting from a probabilistic network repr
esentation, we develop a structure for presenting visual knowledge, an
d techniques for probability propagation, parameter learning and struc
tural improvement. This framework provides an adequate basis for repre
senting uncertain knowledge in computer vision, especially in complex
natural environments. It has been tested in a realistic problem in end
oscopy, performing image interpretation with good results. We consider
that it can be applied in other domains, providing a coherent basis f
or developing knowledge-based vision systems.