Abstract

Recovering useful descriptions of images is a central problem for vision. The minimal-description-length (MDL) criterion is a practical way to evaluate the relative usefulness of a set of surface descriptions as an overall approximation of the image. This criterion minimizes the residual error of the approximation as well as the cost of specifying the individual descriptions. Each individual description ideally models a single process in the scene and may overlap with several other descriptions to account for a portion of the image. The MDL criterion is equivalent to the Bayesian MAP estimation or Gestalt notions of simplicity in perceptual organization.We present a system that uses the MDL formulation to recover the set of part-based shape descriptions that best describes a given image. Our implementation on a Connection Machine uses parallel robust estimation to generate hypotheses of part shape, followed by a Hopfield-Tank network optimization to find the subset that minimally encodes the image.

© 1990 Optical Society of America