Entropy Masking

Abstract

Current vision theory admits two principal types of visual masking: contrast masking and noise masking. The former is thought to operate via mechanisms such as a nonlinear transducer, contrast gain control, and lateral inhibitory influences. The latter is understood to be due to the randomness of the target + mask combination. Here I propose a new category: entropy masking. Entropy masking is evident when the mask is a complex but deterministic signal. Because the mask is deterministic, and because experiments demonstrate that the masking power of such signals is much greater than predicted by contrast gain control, entropy masking cannot be due to either contrast masking or noise masking. Entropy masking occurs because the observer cannot discount the background, as supposed by ideal-observer theory, and hence must learn the target + mask signal. The ease with which this is done is a function of the entropy of both target and mask. Experiments that illustrate these general principles are described.