Abstract
As part of a long term goal to understand visual search in natural scenes, we are currently investigating factors that limit target detection in natural backgrounds. Here we investigate how the statistical properties of natural images affect the behavior of a rational detector. We use a large number of natural image patches (∼300000) to train a PCA-based, k-nearest neighbor (NN) algorithm, where the task is to classify stimulus patches as either a natural patch + gabor target or a natural patch alone. We have some indirect evidence that this algorithm achieves near-optimal performance. First, the NN algorithm reaches near-optimal performance when detecting targets embedded in 1/f noise patches, and second, the detection accuracy of both the natural and 1/f noise NN classifiers asymptotes when the number of patches reaches 100000. In agreement with ideal-observer theory, we find that for gabor targets in 1/f noise (with no signal uncertainty), d' is a linear function of target contrast (Weibull slope parameter of 1.4). On the other hand, for gabor targets in natural scenes we find a change in psychometric function shape—d' shows a slight decelerating non-linearity as a function of target contrast (Weibull slope parameter of 1.0–1.2), which is opposite to the effect of target uncertainty. We are currently using a similar NN simulation technique to evaluate the effects of retinal filtering (eccentricity effects) and target uncertainty on the shape of the psychometric function for optimal detection in natural images.
Supported by NIH grant EY02388.