Abstract
Classification image analysis has proven to be a valuable tool for revealing features used to perform visual tasks in noise. We use this methodology to investigate how the amount of noise in a stimulus influences detection mechanisms. Experiments used to test models of detection in noise span contrast levels ranging from less than 1% to well over 10% with a corresponding range of noise spectral densities. Furthermore, experiments that vary the spectral density of a stimulus have been used as a way to determine equivalent internal noise power. The generality of these approaches depends on the assumption of a common detection strategy. We test this assumption by measuring classification images for two-alternative forced-choice (2AFC) detection of a small Gaussian target (width 8.5 min) embedded in static noise at spectral densities that range from 0.27·10−6 deg2 to 6.7·10−6deg2. Signal contrast was manipulated to maintain a performance level of approximately 85% correct. Classification images were computed from 2000 2AFC trials in low- and high-noise experiments. A spatial frequency analysis was performed by converting the spatial classification images to the frequency domain using the real part of the Discrete Fourier Transform, and averaging over radial bands of spatial frequency. The high-contrast noise classification images show a mild peak at 2–3 cyc/deg before falling off with the frequency spectrum of the Gaussian. The low-contrast noise classification images show a stronger peak at 4 cyc/deg. The different classification images we observe suggest that mechanisms of detection change with the spectral density of a stimulus.