What are the basic image attributes sensed by human vision? This fundamental question has proved difficult to answer experimentally. We introduce a novel psychophysical method that provides leverage for addressing this question in the context of visual texture perception. On each trial, the participant sees a brief display comprising a randomly positioned set of circular apertures, each filled with texture. Some of the apertures contain a "distractor texture" D; others contain "target texture" T. The task of the participant is to mouse-click the centroid of the set of T-apertures, while ignoring the D-apertures. Suppose the participant performs this task using a separable linear computation: (1) computing a set of neural images corresponding to preattentive mechanisms, M_k; (2) combining these images into a weighted average image S (with nonnegative weights w_k); and (3) extracting the centroid of the resulting image. An ideal observer, that aims to minimize the Euclidean distance of the response from the target centroid, should choose the w_k to maximize S(T)/S(D), where S(T) and S(D) are the weighted averages of mechanism responses to textures T and D, respectively. This ratio is maximized by assigning all the weight to the single mechanism M_k for which M_k(T)/M_k(D) is largest. Thus, if a participant performs as well as possible in the centroid task, the resulting behavior reflects use of a single mechanism. We apply this method to white noise textures, varying the distributions of grayscale pixel values characterizing D and T in different conditions. Results implicate (1) a "blackshot" mechanism, sharply tuned to the blackest pixels; (2) a "dark-gray" mechanism with maximal sensitivity for pixels between black and mid-gray, (3) a "down-ramped" mechanism whose sensitivity is maximal for black and decreases quasi-linearly with luminance, and (4) a complementary up-ramped" mechanism whose sensitivity increases linearly with luminance, with maximum sensitivity to white.