ABSTRACT
A framework for machine vision is sketched in order to discuss current research problems in the evaluation of digitized images. The recognition and description of objects, of their shape, texture, and other properties, as well as a description of their static spatial relations with respect to each other and to the recording camera, all require complex system-internal representations.
This framework is extended to discuss the investigation of image sequences which permit temporal developments within the field of view of a recording camera to be captured. The temporal variation of image intensities can be exploited to infer temporal variations of geometrical relations in space. Although this raises a number of very interesting research problems, this contribution emphasises a different research avenue: linking the results to be extracted from an image sequence to descriptions at the level of natural language sentences. In the field of computational linguistics, considerable experience has been accumulated in the construction of internal representations for natural language statements about the world. Tapping this experience for research into the evaluation of image sequences is expected to be of advantage.
In this paper we describe a system aimed at providing software support for the process of knowledge acquisition. Such support comprises a workbench incorporating a number of knowledge acquisition tools. These include knowledge elicitation techniques such as sorting and rating methods, together with machine learning techniques. The paper discusses the various problems raised by this work. These include: defining an adequate view of the general acquisition process, developing an appropriate implementation architecture, directing knowledge acquisition via knowledge level models and producing a sufficiently powerful representation language to integrate the results of acquisition. Finally, we describe the limitations of our current system and propose future developments in our work.
