Abstract
We developed a model of figure-ground organization based on neurophysiological evidence for border ownership-coding in primate visual cortex (Zhou et al., J Neurosci 20:6594, 2000). Figure-ground organization is conceived as a bottom-up process generating a visual data structure that serves top-down mechanisms of information selection. The proposed model consists of a layer of edge detector cells E in which each position and orientation is represented by a pair of cells corresponding to the two sides of ownership, and a layer of grouping cells G that sum the signals of E cells according to roughly circular templates, and, through feedback, control the gain of the E cells. Each E cell excites only G cells on one side of its receptive field (the ownership side), and the two members of an E pair inhibit each other. T junctions are represented by a biased pair of E cells. The G-cells are thought to provide the handles for top-down selection mechanisms. In a previous version (Schuetze et al., VSS 2003) we assumed that the G cell feedback enhances the gain of the corresponding E cells. However, as more template sizes were included to achieve scale invariance, the positive feedback scheme made the model increasingly unstable. We have now replaced the gain enhancement in E cells by divisive inhibition of their antagonistic partner cells. Each E cell activates G cells with tangent templates of all sizes, and the partner E cell is inhibited by the pooled activity of these G cells. The new model yields scale-invariant, correct border-ownership assignments for squares, overlapping squares, and C-shaped figures. In addition to ensuring stability in the model the divisive inhibition mechanism also accounts naturally for the experimental observation that surround stimulation has a multiplicative effect on the classical receptive field response.
Supported by NIH grant EY 02966