Knowledge Acquisition for Image Understanding

Image understanding often requires extensive background knowledge. For example, to the lay person, an x-ray image of the human cerebral vasculature can be almost meaningless. However, a trained radiologist, with a knowledge of anatomy and years of experience, can usually quickly detect anomalies in the blood vessels in the x-ray image. The problem addressed in this paper is how the background knowledge for a complex image processing task can be acquired. We discuss how relational machine learning methods can be used to automatically build rules for classifying abnormalities and we speculate on how interactive knowledge acquisition tools, such as ripple-down rules, may be used to refine to the knowledge.