Complexities of complex contrast

Abstract

For the visual system, luminance contrast is a fundamental property of images, and is one of the main inputs of any simulation of visual processing. Many models intended to evaluate visual properties such as image discriminability compute perceived contrast by using contrast sensitivity functions derived from studies of human spatial vision. Such use is of questionable validity even for such applications (i.e. full-reference image quality metrics), but it is usually inappropriate for no-reference image quality measures. In this paper, we outline why the contrast sensitivity functions commonly used are not appropriate in such applications, and why weighting suprathreshold contrasts by any sensitivity function can be misleading. We propose that rather than weighting image contrasts (or contrast differences) by some assumed sensitivity function, it would be more useful for most purposes requiring estimates of perceived contrast or quality to develop an estimate of efficiency: how much of an image is making it past the relevant thresholds. 1. CONTRAST 1.1 Primacy For the visual system, luminance contrast is the fundamental carrier of information about images. Motion is perceived through temporal changes in luminance contrast; the most initial sensations of depth are formed from binocular combinations of monocularly sensed contrasts; chromatic variation is almost always correlated with luminance changes. All of these qualities (motion, depth, color) are important parts of normal visual experience, and thus of any full-quality representation of it; but a monochromatic, cyclopean still-image is a perfectly acceptable visual representation of a scene. Here we argue that in estimating the visual quality of an image, contrast thresholds are of principal importance; perceived (suprathreshold) contrast magnitudes although noticeable in side-by-side comparison are relatively less important; and that the specific sensitivity functions commonly used in standard practice to estimate perceived contrast and quality may be misapplied or inappropriate. 1.2 Measurement Given the primary importance to vision of luminance contrast, it is of great practical and theoretical importance to have operational measures of it (1). The simplest summary measures will usually fail in characterizing the apparent contrast of an image, and thus are not used except for the simplest of patterns. Michelson contrast, the absolute range of luminances in a pattern, is thus not widely used except for periodic grating patterns - it errs by ignoring too much of an image's spatial variation. A much more common measure of a complex image's contrast is the standard deviation of luminances in an image (RMS contrast), a measure of the average deviation in luminance from the image mean over a specified spatial area. This measure is less susceptible to extreme values in an image, and thus tracks better with perceptual appearance of image contrast. Still, RMS contrast is a relatively poor predictor of perceived contrast - it errs by equally weighting all of the image's spatial variation and has been shown repeatedly to fail in predicting perceived image quality. 1.3 Perception The visual system responds to images through a system of overlapping neural networks, repeated across the visual field, which are sensitive to different spatial scales, and perceived contrast is related to the response of these networks. The