Purchase this article with an account.
HR Wilson, F Wilkinson, G Loffler; The Role of Principal Components in Synthetic Face Discrimination . Invest. Ophthalmol. Vis. Sci. 2002;43(13):4791.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Purpose:Although principal component analysis (PCA) has become a popular technique for the analysis of human faces, there is as yet no experimental evidence that PCA is actually used by the visual system in facial analysis. Accordingly, our goal was to experimentally test the hypothesis that the visual system employs PCA in a face discrimination task. Methods:Thresholds were measured for synthetic faces, each derived from a face photograph by digitization of 37 salient points. All synthetic faces are bandpass filtered with a circularly symmetric DOG filter having a 2.0 octave bandwidth and a peak frequency of 10 cycles/face width. Thresholds were measured for discriminating individual synthetic faces from the mean for each gender. In addition, a PCA was performed on the population of 80 synthetic faces, and thresholds were measured for discriminating PCA faces from the mean for each gender. Results:Thresholds for synthetic faces averaged 7% geometric variation from the mean face for both front and 20° side views. Thresholds for PCA faces, however, averaged about 2.9% geometric variation, 2.4 times better than for the original synthetic faces. Calculations indicate that the information used for discrimination must be extracted from a global spatial analysis of the stimuli. Conclusion:Our data are consistent with a PCA population code for the geometric information contained in synthetic faces. The higher thresholds for synthetic faces derived from individual faces can be predicted on the assumption that discrimination occurs when the largest principal component in any given face reaches threshold.
This PDF is available to Subscribers Only