Parameter Networks: Towards a Theory of Low-Level Vision

One of the most fundamental problems in vision is segmentation; the way in which parts of an image are perceived as a meaningful whole. Recent work has shown how to calculate images of physical parameters from raw intensity data. Such images are known as intrinsic images, and examples are images of velocity (optical flow), surface orientation, occluding contour, and disparity. While intrinsic images are not segmented, they are distinctly easier to segment than the original intensity image. Segments can be detected by a general Hough transform technique. Networks of feature parameters are appended to the intrinsic image organization. Then the intrinsic image points are mapped into these networks. This mapping will be many-to-one onto parameter values that represent segments. This basic method is extended into a general representation and control technique with the addition of three main ideas: abstraction levels; sequential search; and tight counting These ideas are a nucleus of a connectionist theory of low 'eve and m'ermediate-level vision. This theory explains segmentation in terms of massively parallel cooperative computation among intrinsic images and a set of parameter spaces at different levels of abstraction. The preparation of this paper was supported in part by the Defense Advanced Research Projects Agency, monitored by the ONR, under Contracts N00014 78-C0164 & N000J4-80 -C-0197.