Knowledge and the visual process: Content, form and use

摘要

A knowledge based system for the analysis of imagery clearly requires large amounts of domain specific knowledge and also requires a recognition control scheme that will manipulate this knowledge in order to interpret the imagery that represents various scenes of the domain. Many current systems indeed satisfy this statement. In addition, however, they all contain modules that access the actual image data and process this data. Typically, the methodologies for the image specific aspects and the domain specific aspects are separate yet interact, and the representational formalisms and control schemes for these two tasks are not related.This paper will attempt, by overviewing a current hypothesis of the kinds of knowledge required for general purpose vision and the current representational tools available, to reconcile the “low” and “high” levels of knowledge based vision systems and to propose a set of uniform representational tools. The discussion will be at the conceptual level and not at the implementational level. Pointers to current computer vision schemes that are relevant to the discussion will be given. Several good surveys and discussions of requirements of vision systems can be found in Nevatia,(1) Nagel,(2) Hanson and Riseman,(3) Barrow,(4) Weszka,(5) Reddy,(6) and Kanade.(7)