부산대학교 도서관

Stereocorrespondence is one of the classical problems in machine vision. The problem concerns the determination of corresponding points in a pair of images taken from two vantage points. Correspondence is difficult to find when the two images contain many identical features. The author describes an improved stereocorrespondence algorithm based on a combination of computational, psychophysical, and neurophysiological data, and the implementation of this algorithm in an analog VLSI chip that computes stereocorrespondence from one-dimensional input patterns in real time. This algorithm is naturally robust against random transistor mismatch, which, from the point of view of the circuit, is similar to false correspondences in the visual data. This noise-suppressing property makes the stereocorrespondence circuit architecture generally useful for analog VLSI implementations.ETX