Local statistics in natural scenes predict the saliency of synthetic textures

Abstract

AbstractTo represent behaviorally relevant information efficiently, neural circuits in sensory periphery are matched to the first and second-order statistical structure of natural inputs. The retina, for example, removes stimulus components that are predictable (and therefore uninformative), and transmits what is unpredictable (and therefore informative). Here we show that this efficient coding principle applies to complex aspects of natural scenes, and to central visual processing (Tkačik et al, PNAS 2010).The aspect of natural scenes that we examine in detail is texture in image patches. Texture is determined partly by the distribution of light intensities and partly by the spatial organization of light across pixels. To characterize these high-order natural image statistics we use two complementary dimensionality-reduction approaches. For intensity distributions, we analyze spatial variations in local intensity histograms. For local spatial organization of light, we analyze fourth-order correlations of nearby pixels. These approaches robustly classify image statistics according to their informativeness about the local structure of natural images.Remarkably, in both cases, we find that the distinction between informative vs. uninformative high-order statistics corresponds closely to the perceptual sensitivities of the human visual system documented in psychophysical measurements. For intensity distributions, the three most informative histogram statistics of natural images correspond to the three mechanisms that account for perception of spatially unstructured (“independent, identically-distributed”) artificial textures, namely, mean, variance, and a quantity known as “blackshot”. For spatial organization, we find that fourth-order spatial correlations that are informative correspond to those fourth-order correlations that are visually salient. Moreover, sensitivity to the latter high-order correlations is known to arise in visual cortex. Our results suggest that the principle of efficient coding also shapes later stages of sensory processing that are sensitive to high-order image statistics.