On the need for integrated approaches to image understanding

Abstract

Knowledge‐Based (K‐B) approaches to Computer Vision seem to have come to a stalemate, due to difficulties inherent in performing crucial activities, e.g., autonomous refinement of the embedded high‐level knowledge and time‐consuming low‐level processing. In the paper, the problem is related to a broader class of issues, and the conclusion is drawn that no individual method can be regarded as a unique and comprehensive paradigm for Computer Vision. A possible solution to this problem might lie in combining different methodologies to enhance mutual advantages and to balance individual shortcomings; K‐B vision approaches are reviewed here from this integration‐oriented perspective. Higher‐level integration should aim at improving 'intelligent' activities of K‐B systems; Machine Learning approaches to Vision are discussed from this point of view, Lower‐level integration should provide K‐B systems with the fast number‐crunching capability and robustness required; the use of neural networks and of associative memories is evaluated for this purpose. All such methodologies are reviewed, taking into account both their intrinsic peculiarities and the specific problems that may arise from coupling them with K‐B vision systems. The paper aims at providing an updated and critical picture of some lines of research that are currently being pursued to achieve an efficient integration of K‐B techniques into complex vision systems.