Abstract
The statistical regularities of natural signals are a starting point for understanding the characteristics of early visual processing, e.g. the center-surround architecture of retinal ganglion cell receptive fields. Can this matching between natural signal statistics and neural processing mechanisms be extended beyond the sensory periphery? Our recent work (Tkacik et al., 2010) showed that human sensitivity to isodipole (fourth-order correlated) synthetic textures, known to arise in cortex, is closely related to the structure of fourth-order spatial correlations in natural scenes. This leads us to propose a specific organizing principle: The perceptual salience of visual textures increases with the variance (i.e. unpredictability) of the corresponding correlations over the ensemble of natural scenes. To test this idea we focused on local image statistics: correlations between two, three, and four adjacent pixels within a 2x2 square. For binarized images, there are four pairwise correlations -- vertical (betaV), horizontal (betaH) and diagonal (beta,beta/) -- four third-order correlations (theta1,2,3,4) and one fourth-order correlation (alpha). We measured these correlations in a large ensemble of image patches taken from the UPenn Natural-Image Database. The variance in these correlations over the ensemble was robustly ordered (increasingly) as: Var(theta1,2,3,4),Var(alpha),Var(betaDiag), Var(betaV,H). Thus, our broad hypothesis predicted the same ordering of perceptual salience of artificial textures with correlations of different types. Furthermore, our hypothesis predicts that the principal components of the covariance of these correlations in natural scenes match the principal components of observers' perceptual sensitivities to synthetic textures. These predictions are confirmed in psychophysical experiments: observers' ability to use image statistics to segment artificial visual textures conformed to the ordering of their variances in natural images, and observers' principal axes of sensitivity lie near the principal axes of covariances of the image statistics. Our results suggest central neural mechanisms are efficiently tuned to the statistics of natural scenes.
Meeting abstract presented at VSS 2013