Evidential Knowledge-Based Computer Vision.

摘要

It has been argued that knowledge-based systems (KBS) must reason from evidential information - i.e., information that is to some degree uncertain, imprecise, and occasionally inaccurate. This is no less true of KBS that operate in the domain of computer-based image interpretation. Recent research has suggested that the work of Dempster and Shafer (DS) provides a viable alternative to Bayesian-based techniques for reasoning from evidential information. In this paper, we discuss some of the differences between the DS theory and some popular Bayesian-based approaches to effecting the reasoning task. We then discuss some work on integrating the DS theory into a knowledge- based high-level computer vision system in order to examine various aspects of this new technology that have not been explored to date. Results from a large number of image interpretation experiments will be presented. These results suggest that a KBS's performance improves substantially when it exploits various features of the DS theory that are not readily available in pure Bayesian-based approaches.