|
|
|
“Intelligent Interpretation of Aerial Images”
by V. Venkateswar and R. Chellappa
March 1989
The task of interpreting aerial images can be divided into two sub-tasks - low-level and high-level. The purpose of the low-level module is to abstract some primitives (e.g.: lines, regions) from the raw image data to be input to the high-level module. The high-level module then attempts to synthesize these input primitives into objects using various sources of knowledge. However, the output from the lower level module is frequently ambiguous, fragmented and insufficient. Consequently, there is a search space involved as in many other artificial intelligence problems. Purely algorithmic procedures, that make ad hoc choices when faced with alternatives, are bound to have limited practical use. Some kind of intelligent search mechanism needs to be employed. At the Signal and Image Processing Institute, USC, we are implementing a system for intelligent interpretation of aerial images. The lower-level module is a new line detector. An ATMS-type multiple context truth maintenance system (MCTMS) oversees search in multiple contexts. This search scheme is well suited to aerial image interpretation. Alternative paths can be simultaneously explored and competing contexts can be directly compared. The process of object synthesis is carried out in a hierarchical manner. An object is composed of faces, faces are formed by edge rings, edge rings are formed by edges, which are defined by their end vertices. Each element in this hierarchy is representation unit - a frame. Relations between these elements are also represented as frames. The frame paradigm permits us to perform expectation driven processing and also facilitates object oriented programming. The domain of interpretation is currently restricted to buildings that are compositions of rectangular parallelepipeds. Currently our system is only partially complete. The system is capable of composing edge rings and can detect closed edge rings.