Abstract
Although the issue of how humans identify faces has received much research attention, the question of face detection (‘is this a face?’) remains largely unexplored. In this work, we investigate whether the representations subserving face-detection emphasize the holistic facial structure or individual facial components. Our work has two parts — 1. experimental characterization of human face/non-face discrimination performance, and 2. comparison of human data with those from computational systems for face detection. The face stimulus-set comprised clear facial images as well as several transformations thereof, including inversion, Gaussian blurring, contrast negation, pixel noise addition, and component removal. These transformations were designed to differentially impact the holistic and component-based processing strategies. The non-face stimuli comprised false-alarms of notable machine-based face-detection systems and patterns with similar Fourier spectra as faces. We found that human performance is very robust to facial image degradations: Our results so far show that the removal of single components, vertical inversion and contrast negation do not lead to statistically significant decrements in classification performance. However, Gaussian blurring and pixel noise addition result in marked deficits. Furthermore, these detection results show interesting differences relative to those reported for face-identification (for instance, the differential consequences of inversion and negation on the two tasks). We shall also describe simulations with two kinds of Support Vector Machines classifiers: one set up as a hierarchy with the first stage detecting facial components (i.e. piece-meal processing), and the other set up as a single template classifier (i.e. holistic processing). The goal is to determine which computational strategy produces results more congruent with the experimental data from humans.