Abstract
Most studies of detection in complex backgrounds have measured and modeled human performance for statistically uniform (stationary) backgrounds. However, natural and medical images have statistical properties that vary over space. We measured the detection psychometric function for various target shapes presented in Gaussian 1/f noise backgrounds that were statistically uniform over space and that modulated in contrast over space. We used 1/f noise backgrounds because they have the approximate amplitude spectrum of natural images. We compared human performance with the ideal observer and five sub-optimal observers. Each model observer had only one free parameter: an overall efficiency scalar. We find that the pattern of human thresholds is not consistent with the ideal observer, but is quantitatively consistent with a suboptimal observer that performs partial whitening in spatial frequency (consistent with the average human CSF) and optimal whitening (reliability weighting) in space, and has a small level of intrinsic position uncertainty, which was estimated independently in a separate experiment. It is likely that the three factors in this suboptimal observer will be important for explaining and predicting detection performance in a wide range of natural and medical images.