Effect of higher-order statistics of images on signal detection performance of human observers

Abstract

Recent work has shown that a linear discriminant model derived from the work of Harold Hotelling can account for a significant body of psychophysical data.1-3 This model utilizes only the first- and second-order statistics of the image and is insensitive to the shape of the grey level histogram. A natural question is whether the human observer is also insensitive to the shape of the histogram or whether human observer performance is sensitive to higher-order statistics in a image. To answer this question, a psychophysical study was conducted. The images viewed by human observers were simulated ones with inhomogeneous, random backgrounds, and Poisson noise. The mean, variance, and autocorrelation function of the images were controlled to be the same for all images. The grey level histograms of half of the images were designed to be distinctly non-Gaussian, while the other half had Gaussian histograms. Thus the first- and second-order statistics were constant, and only higher-orders were variable. Our results indicate that human detectability is independent of the shape of the grey level histograms, and that the Hotelling observer remains a good predictor of human performance in spite of this variability.