One of the central issues in vision is how to represent and use knowledge relevant to understanding the image. Partly because vision is so difficult. and partly because even the cheapest solutions can still be so useful. Approaches to vision problems have had a tendency to be ad hoc and heunstic. Recently, however. new thrusts in computer vision are emerging. most notably in the Image Understanding community, that try to pursue more systematic and computational approaches. [l] This article attempts to introduce such new Image Understanding approaches to vision. It first presents the author's view of structure of vision: what types of information must be dealt with and what levels of knowledge are involved in transforming one type of information into another. Then. representative research progress in Image Understanding is reviewed which addresses use and representation of knowledge in vision. Specifically. we will discuss: Formalization of physical knowledge into computational forms Use of 3D models Construction of scenic descriptions from images Organization and control of vision systems