Image reconstruction using natural scene statistics predicts many aspects of lightness perception
The perceived lightness of objects is consistent under huge variations in illumination. This is remarkable because it requires that the visual system take one effect – the amount of light landing on the retina – and disentangle the contribution of two causes: the reflectance of surfaces and the intensity of their illumination. Solution of such an under-constrained problem must therefore involve assumptions, which could be based, for example, on high-level knowledge of objects. Here we propose that such assumptions are largely based on the statistical properties of natural scenes and, specifically, that lightness perception can be posed as a reconstruction problem where the visual system infers the image most likely to have caused the particular outputs over a bank of spatial filters. Reconstruction then involves two assumptions. First, that images exhibit scale-invariance - filters at different spatial frequencies contribute to a a simple linear relationship to one another on log-log axes, (a.k.a. 1/f statistics). Second, that illumination tends to vary slowly over scenes (compared to features such as edges), so that it’s effect may be minimised by switching out a subset of the low-frequency filters (local gain control). We show that imposition of these two properties on images can both adequately reconstruct natural scenes and account for the presence and magnitude of a wide range of lightness illusions.