Abstract

Algorithms for analyzing visual scenes often depend on isolating the relevant visual cues and extracting the cues from images. For example, disparity is a cue for binocular depth estimation, and the depth is extracted by solving the correspondence problem. Problems arise, however, when more than one cue is relevant and the complexity of their interactions makes it difficult to analyze all possible contingencies. A new technique has recently been developed for neural network architectures in which the relevant cues and their contingencies are learned from examples. Images and their correct interpretations are supplied to the learning algorithm, which constructs feature detectors and parameter filters to solve the problem. The application of neural network learning algorithms to two problems in vision is presented: The extraction of shape parameters from shaded images of surfaces and the estimation of depth from binocular images. The properties of processing units in the trained networks are compared with those of neurons in mammalian visual cortex.

© 1988 Optical Society of America