Abstract
Previous work (Nandy & Tjan, 2008) has shown that human observers are able to linearly integrate spatial-frequency information within letter stimuli. The purpose of this study was to determine if this ability is preserved for the task of face identification. Our stimuli consisted of 1074 standardized face images of celebrities. Stimuli were 4° in width. Each image was filtered using three different raised cosine log filters with a bandwidth of 1 octave. Two filters had center frequencies of 8 cycles/face and 32 cycles/face. These frequencies were chosen because they produced stimuli that yielded similar contrast thresholds for identification. The third filter was created by adding the first two filters. We measured contrast thresholds for identifying face images for each filter condition in seven human observers, and also for a white-noise limited ideal observer. For comparison, we measured contrast thresholds for identifying letter stimuli (x-height = 0.25°). Letters were filtered with similar filters centered at 1.35 cycles/letter and 5.4 cycles/letter as well the combination of the two. To quantify linearity, we adopted the same metric as Nandy & Tjan – the integration index, defined as the ratio between the squared contrast sensitivity of the composite and the sum of squared contrast sensitivities of component images, where a value of 1 implies linearity. For human observers, the integration index for identifying face and letter stimuli averaged 3.69±10.94 and 0.93±0.22, respectively. In comparison, ideal observer analyses yielded integration indices of 0.17±0.02 and 0.93±0.06 for faces and letters respectively. While the finding of our letter experiment replicates that of Nandy & Tjan, the results of our face identification experiment however point to a non-linear integration of spatial-frequency information. The ideal observer analysis suggests that our results could be explained by the differences in the distribution of spatial-frequency information between face and letter stimuli.
Meeting abstract presented at VSS 2012