The “super-importance of hue” in psychophysics, physiology, and AI

Abstract

Hue is thought to play a more important role for color discrimination than chroma. However, whether this is true appears to depend on the region of color space. In physiologically-based color opponent spaces, previous literature has shown that for orangish and bluish colors, hue thresholds are much smaller than chroma. But for purplish and greenish colors, hue and chroma thresholds are nearly equal. It is unclear why the visual system prioritizes some mechanisms over others. We first present neural evidence of this effect: we took magnetoencephalography measurements while subjects performed a simple discrimination task with either purplish or orangish uniform discs. We found greater amplitude modulation – indicating better discriminability – when the orangish stimuli differed in hue compared to chroma, and compared to both hue and chroma for purplish stimuli. Behaviorally, we explored whether the psychophysical differences would also arise in color discrimination of single-hue rendered objects, which elicit a distribution of points in color space. We found thresholds were merely elevated for the rendered objects compared to single patches of light. Hue superiority was present only for orangish colors, not purplish. We then analyzed multiple image databases to see whether the color distribution of objects found in the environment was biased and found that they overwhelmingly plot in the orangish regions of color space. We used the linear probe technique to interpret the internal representation of several deep neural networks trained on such biased image sets with high-level visual tasks such as object recognition and text-image pairing. The pattern of thresholds estimated from the networks, in particular the hue-chroma asymmetry, was similar to humans. We conclude that hue is indeed of superior importance for color discrimination, and that the peculiarities of this are shaped by the color statistics of our environment.