Abstract
Purpose: Previous studies have proposed estimation procedures for classification images in white and correlated noise (Abbey & Eckstein 2006). Here, we investigate methods to estimate classification images (linear template) for search of targets embedded in natural images. Methods: Observers searched for an additive Gaussian luminance target in one of four locations (4 alternative forced choice) within 3000 calibrated natural scenes (van Hateren & van der Schaaf, 1998). We compute classification images using various methods including genetic algorithms (GA; Castella et al., 2007), support vector machines (SVM; Jakel et al., 2009) and weighted averaging of prewhitened noise fields (Abbey & Eckstein, 2002). All methods relied on a limited set of Gabor basis functions. We compare human classification images to optimal linear templates also estimated using GA and SVM. Results: GA and SVM methods result in similar estimates of the optimal linear templates. Average observer performance was higher than the estimated human classification images and optimal linear templates. For all three observers, the estimated linear templates were similar to the target but contain inhibitory surroundings. Conclusions: We extend previous classification image methods to search for a target embedded in natural scenes and explore computational procedures for reliable estimation. The presence of inhibitory surrounds in human classification images reflects a strategy to optimize detection of targets in natural scenes. The superior human performance relative to the optimal linear template suggests that humans are able to use additional higher-order information to detect targets in natural scenes.