Citation: Neural Networks, 2, 29-51
Abstract: A neural network model of multiple-scale binocular fusion and rivalry in visual cortex is described and simulated on the computer. The model consists of three parts: a distributed spatial representation of binocular input patterns among simple cells that are organized into ocular dominance columns; an adaptive filter from simple cells to complex cells; and a nonlinear on-center off-surround shunting feedback network that joins together the complex cells. This data structure generates complex cell receptive fields which multiplex input position, orientation, spatial frequency, positional disparity, and orientational disparity, and which are insensitive to direction-of-contrast in the image. Multiple copies of this circuit are replicated in the model using receptive fields of different sizes. Within each such circuit, the simple cell and complex cell receptive field sizes covary. Together these circuits define a self-similar multiple-scale network. The self-similarity property across spatial scales enables the network to exhibit a size-disparity correlation, whereby simultaneous binocular fusion and rivalry can occur among the spatial scales corresponding to a given retinal region. It is shown that a laminar organization of the model interactions among the complex cells gives rise to conceptually simple growth rules for intercellular connections. The output patterns of the model complex cells are designed to feed into the model hypercomplex cells at the first competitive stage of a Boundary Contour System network, where they trigger a process of multiple-scale emergent binocular boundary segmentation. The modeling results are compared with psychophysical data about binocular fusion and rivalry, as well as with the cepstrum stereo model of Yeshurun and Schwartz. The results indicate that analogous self-similar multiple-scale neural networks may be used to carry out data fusion of many other types of spatially organized data structures.