Abstract
Two experiments investigated the perceptual learning of spatial frequency content in grey-scale images of human faces. In Experiment 1, facial stimuli were filtered by either low- or high-pass ideal filters to create two conditions: Low (4–15 cycles/image) and High (16–63 cycles/image). Stimulus contrast was varied to measure identification thresholds in a 10–AFC matching task. Perceptual learning was found to be frequency-specific: observers trained to identify faces in the Low condition did not show improved performance when tested in the High condition, and vice versa. This result is consistent with studies that show that a certain degree of spatial frequency overlap is required for accurate matching of filtered faces (Liu et al., 2000). To investigate the effects of spatial frequency overlap more directly, observers in Experiment 2 were required to match filtered facial stimuli to All-Pass facial images. Faces were filtered to create 3 conditions, each with a 2-octave bandwidth: Low (2–7 cycles/image), Medium (8–31 cycles/image), and High (32–127 cycles/image). In a 4-AFC matching task, observers were able to accurately match the Medium faces to the correct All-Pass counterparts, but performed much worse in the Low and High conditions. An ideal observer analysis showed that the results could not be explained by differential amounts of information within different spatial frequency bands, and additional control experiments showed that the pattern of human performance was not due to the shape of the contrast sensitivity function. These results support the idea that observers can learn to identify faces when there is some overlap in spatial frequency content, but that the degree of overlap does not constrain performance per se. Instead, the middle band of frequencies plays a special role in face identification.