Abstract

Most machine vision systems compute curvature in three steps: (1) blob or contour extraction; (2) tangent computation along an edge list; and (3) tangent differentiation to obtain curvature. Such a procedure requires choosing the spatial scale for these different operators (blob, tangent, and curvature). Customarily, a range of 2-D masks and 1-D curvature operators is used to capture as much information as possible. In contrast, our recent psychophysical data suggest that the human visual system computes curvature using only one spatial scale, namely, the finest, except when pressed to perform on fuzzy contours. Furthermore, two different types of mechanism are required to cover the range of curvatures discriminated. For low curvatures, the mechanism resembles the cocircularity comparison of bar masks proposed by Parent and Zucker (1985). For high curvatures, our visual system uses a single receptive field operator analogous to that proposed by Koenderink and van Doom (1986). (12 min)

© 1987 Optical Society of America