Abstract
Performance detecting a signal typically degrades in the presence of visual stochastic noise and deterministic masks. Less is known about how human performance is affected by the presence of probabilistic masks that vary across the possible signal locations. Do human observers lump these visual patterns with the noise or do they use knowledge about the masks' shapes to discount them? To investigate this question, we measured human performance detecting a circular Gaussian shaped signal embedded in one of four locations (4 AFC) in two conditions: a) white Gaussian noise; b) white Gaussian noise plus four additive probabilistic masks (one at each location) randomly and independently sampled with replacement for each location and trial from a collection of 720 different elongated Gaussian masks with differing shape, size, contrast and polarity. Human performance was compared to a model that used information about the signal but ignored the probabilistic masks (i.e. lumped the masks with the noise; non-prewhitening matched filter), the linear observer leading to maximal d' (i.e. using the changes in image variance and covariance introduced by the masks), and the optimal Bayesian observer that uses full knowledge about the signal and the masks and a non-linear decision rule. When internal noise is added to the two linear models so that they match human performance in the white noise only condition, then human performance in the probabilistic mask condition exceeds performance of the two linear models. This suggests that humans discount the masks based on knowledge about their shapes and non-linear decision rules; although very suboptimally compared to the optimal observer.