Classification images (CI’s) are a psychophysical probe of the computations underlying visual perception. However, application of CI’s to texture discrimination is not straightforward, since standard reverse-correlation will not capture the contribution of second- or higher- order image statistics.
Five subjects identified the location of a 16 x 64-pixel texture-defined target within a 64 x 64-pixel background array. Target and background were chosen from a two-dimensional space of binary Markov random field textures, parameterized by their mean luminance and a 2x2 fourth-order correlation. Stimuli spanned the range of performance from near threshold to near ceiling. 4320 trials per subject were collected.
CI’s were determined after preprocessing stimuli to create “derived images” representing pixel-by-pixel estimates of luminance or higher-order statistics. Reverse correlation yielded CI’s that identified the footprint of the target but did not reveal internal structure. However, CI’s extracted by regression combined with regularization identified features not seen in the reverse correlation CI’s: an accentuation of the contribution of luminance statistics, but not fourth- order statistics, near the target edge. Thus, CI’s reflecting nonlinear processes may be readily obtained via analysis of appropriate derived images, and regularization techniques may provide insights beyond those apparent from standard reverse correlation maps.