Reasoning About Images: Application to Aerial Image Understanding

Abstract

Image Understanding (I.U.) shares with Artificial Intelligence the need for mechanisms of using knowledge to control computation. In I.U., this knowledge takes the form of prior knowledge about objects and knowledge gained from doing the computations themselves. This paper presents an approach to the general I.U. problem and a specific program for locating buildings in aerial photographs. Prior knowledge is stored as an Appearance Model, which represents the appearances of possible buildings. A three stage program generates operator and parameter sequences to achieve recognition. Each level uses partial results from below to 1) search parameters for the best match, 2) infer obscuring image conditions and deal with them, and 3) infer high-level conditions such as the presence of occluding objects. This use of partial results is a kind of "meta-knowledge", enabling the program to evaluate its progress and change hypotheses and computations accordingly.