Low level computer vision involves the extraction of photometric
 invariant features such as corners, straight line, and arc segments,
 which represent parts of object boundaries. High level vision concerns
 with matching pre-stored object models to these observed low-level
 image features. The combinatorics of this high level matching process
 grows exponentially with the number of low-level features. Thus, there
 is a need for an intermediate process that is able to select or group
 salient subsets of low-level features without, of course, the
 knowledge of the objects that are present in the image.
 
 The process of perceptual organization that is present in the human
 visual system, which was discovered by the Gestalt psychologists in
 the 1920s, offers an elegant mechanism to effect this intermediate
 level grouping process. However, there are practical problems
 associated with implementing these Gestalt principles of organization,
 as they are called. These principles are essentially qualitative. We
 need to quantify these relations. What are the mechanisms in which
 this can achieved? How would one learn the parameters associated with
 this quantification scheme from examples? In the talk, I will present
 some ideas based on Bayesian networks and Learning Automata that we
 have found to be useful.