A neural architecture of brightness perception: non-linear contrast detection and geometry-driven diffusion

摘要

A neural architecture for brightness perception is constructed in the tradition of filling-in theories. The model is developed to account for a wide variety of difficult data, including the classical phenomenon of Mach bands, low- and high-contrast missing fundamental and non-linear contrast effects associated with sinusoidal luminance waves. The model builds upon previous work by Grossberg and colleagues on filling-in models that predict brightness perception through the interaction of boundary and feature signals. A new interpretation of feature signals through the explicit representation of contrast-driven and luminance-driven information is provided and directly addresses the issue of absolute brightness values. Simulations of the model implement a number of refinements with respect to the previous implementation of Grossberg and Todorović [1] [S. Grossberg, D. Todorović, Neural dynamics of 1-d and 2-d brightness perception: a unified model of classical and recent phenomena, Perception and Psychophysics, 43 (3) (1988) 241–277]. These include: (a) ON and OFF channels with separate fillingin domains; (b) multiple spatial scales; (c) non-linear computations for simple and complex cells; and (d) boundary computations that engage a recurrent competitive circuit. The net effect of mechanisms involved in the computational model accomplish a unique solution for the brightness-from-luminance problem. The two parallel and topographically organized subsystems of a boundary contour (BCS) and feature contour system (FCS) are demonstrated to generate an isomorphic representation of brightness distributions. The activity of the contoursensitive BCS regulates the process of diffusive filling-in. It realizes an adaptive form-sensitive mechanism for the control of lateral spreading of local activation in the diffusion system. It is shown that, under certain stimulus conditions and structure of the input generators to the filling-in processes, the action of BCS/FCS interaction realizes a membrane regularization of the problem of brightness reconstruction. Simulations of the present system of equations account for human perception of a wide variety of stimuli, including the ones studied by Georgeson [2] [M. Georgeson, From filters to features: location, orientation, contrast and blur, Proc. CIBA Symp. on Higher-Order Processing in Vision, London, October 19–21, 1993], whose shallow spatial gradients have posed difficulties to alternative early vision theories. Because boundary signals may undergo reorganization, including long-range grouping before feature diffusion proceeds, the proposed architecture may also serve as an alternative framework for non-linear anisotropic diffusion approaches developed recently for early processes in computer vision.