Novel theory on mach-bands and gradient formation in early vision
Matthias S. Keil, Gabriel Cristóbal, Heiko Neumann · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 2003
Recognizing objects of interest in real-world scenes is a principal computational goal of the primate visual system. This process involves the representation of object surfaces in the first visual areas of the neocortex. Luminance gradients are usually superimposed on object surfaces, what complicates the recovery of the surface reflectance functions on the one hand, but may provide valuable information about surface texture and surface curvature one the other hand. Consequently, there should be a way to recognize and represent luminance gradients independently from object surfaces. However, there is no corresponding theory available up to now. Here we present a two-stage architecture which is compatible with this idea. The first stage involves the detection of luminance gradients in a given intensity image, which are subsequently recovered in the second stage. By means of a novel diffusion paradigm, our architecture is capable of building representations of arbitrary sized luminance gradients from sparse local measurements of gradient evidence. Since our architecture both predicts psychophysical data on Mach bands, and successfully processes real-world scenes, it constitutes a potential computational theory on how luminance gradients are processed and represented in the first visual areas of the brain.