Some Perspectives on Brightness Perception Models through the Study of Brightness Illusions

Abstract

Modeling brightness perception has always been a challenging issue in computer vision for performing real-life, complex visual tasks like object detection, recognition and image analysis under widely varying lighting conditions. Brightness constancy, brightness-contrast, brightness assimilation, transparency etc. are some of the crucial aspects in perceiving brightness that need to be dealt with in order to develop meaningful computational models in machine vision. The fascinating world of brightness illusions provides an important gateway to this study and research. The present thesis attempts to make some humble contributions to this end through both experimental psychophysics as well as from the perspective of computational neuroscience. In this work, we have, first of all, been able to design and establish four new illusory stimuli that expose the limitations of traditional spatial filtering models of vision in general and that of the well-established Oriented Difference of Gaussian (ODOG) filter, in particular. These four new stimuli are all modifications of some of the classical brightness perception illusions. The first is the Mach Band illusion where in our new design, the region of intensity gradient in the image increases linearly in size from zero at the top to its widest at the bottom, so that the bright Mach band is perceived thinner at the top and wider at the bottom diverging outwards like rays of light emanating from the top. Next, is the Hermann grid illusion where we introduced tiny perturbing squares overlapping with each grid square corner, resulting in complete wiping out of the illusory spots in Hermann grid. The third, deals with the sine and square grating stimuli that induce brightness to a foreground uniform test strip. Our design comprises a set of stimuli demonstrating sine to square grating transition unfurling new intriguing problems regarding the mechanism of brightness-contrast phenomenon in the parallel visual pathway and suggest that these two apparently similar effects of brightness induction may occur due to distinctly different mechanisms. The last one is the White's illusion where a simple longitudinal extension of the test patch in our modification demonstrates interesting effects in brightness assimilation and the failure of the ODOG model. We have shown that this limitation of ODOG can be overcome by a new parsimonious spatial filtering model (Difference of Difference of Gaussian or DDOG) that can provide a unified explanation to both brightness contrast and brightness assimilation phenomena, and can also predict many such subtle brightness effects. Two versions of this model viz. the Varying Contrastive Context Filter (VCCF), and the Adaptive Contrastive Context Filter (ACCF) have been proposed, both of which attempt to approximate the Magno (M) and Parvo (P) channels in the central visual pathway. Next, some image processing applications of this new computational model in the domain of denoising with edge preservation are demonstrated especially in comparison to the powerful Bilateral filter. Finally, the limitations of our model and spatial filtering based approaches, in general, have been elucidated and the possible directions of future research in computational theories of brightness perception have been indicated.