Computational theory and neural network model of perceiving shape from shading in monocular depth perception

Proposes a computational theory and a neural network model of perceiving shape from shading in monocular depth perception based on physiological knowledge of visual cerebral cortices. The network estimates the surface orientation of three-dimensional objects from two-dimensional gray-level image data through the interaction between two layers. In the forward connection, approximated inverse optics are calculated by a one-shot algorithm; the surface slant and tilt are estimated from derivatives of image intensity. In the backward connection, optics are calculated; the estimated surface orientation is transformed into the derivatives of image intensity. It was found that the network can estimate the surface orientation of an arbitrary smooth object with a small number of iterations (typically less than ten times)