Abstract
We describe a model for estimating the spatial distribution of visual attention for detecting a target embedded in arbitrarily shaped to-be-attended areas within a field of distracters and non-attended false targets. Expanded from the linear systems model of Gobbel, Tseng, and Sperling (Vision Research, 2004, 44: 1273–1296), the model comprises a spatial frequency modulation transfer function (MTF), an acuity strength map, local compensation for crowding, and a signal-detection decision mechanism. Computation sequence: The pattern of to-be-attended areas undergoes a continuous Fourier transform in order to create an attentional map in the spatial frequency domain that is multiplied by the MTF (attenuating middle and high spatial frequencies), and then inverse Fourier transformed, returning the map to retinal coordinates. The resulting strength map is multiplied by an acuity function, a linear spatial bias function, and a crowding function based on the number of nearest neighbors. This embellished strength map multiplies the input stimuli which consists of a 12×12 array of targets and distractors. Gaussian noise is then added to all locations; the highest valued location is designated as the target. Bootstrap methods are used to derive maximum likelihood estimates (MLE) of model parameters from a set of regular simple stimuli, and then applied to complex 8-block and 16-block patterns of to-be-attended areas (and to the inverse attention patterns). The model rendered parameter-free predictions of search performance in the 72/144 attended locations of the two 8-block and two 16-block patterns; observed data were then evaluated for closeness of MLE fit to model predictions.
This work was supported by the US Air Force Office of Scientific Research, Life Sciences, under grant no. FA9550-04-1-0225.