Abstract
In a biologically motivated recognition system we represent face images as convolution values with a set of multiscale and multiorientation Gabor wavelets (a toy model of V1) at 48 characteristic locations on the face such as corner of the eye, tip of the nose etc. It has been shown previously that human face recognition seems to preserve such V1 similarity (Kalocsai, 1998), but would it also be sensitive to statistics computed upon that representation? The Gabor wavelets described above were ranked according to their discriminative power on face recognition tasks. Then five reconstructed images were created for each face from the wavelet representation: a full reconstructed image; two half reconstructed images: one from the more discriminative (stronger) and one from the less discriminative (weaker) half of the wavelets; and two hybrid reconstructions: one from the stronger half of the wavelets for the target image and from the weaker half for a randomly chosen other image and another one the other way around. The images were adjusted for mean square error to work against the purpose of the study. Subjects performed two forced choice tasks in which they compared the two half reconstructed and the two hybrid reconstructed images to the full reconstructed version. Human observers predominantly choose the face reconstructed from the stronger half of the wavelets to be more similar to the full reconstructed image. The result was the same for the hybrid images where subjects overwhelmingly chose the ‘stronger half from same and weaker half from another’ hybrid image to be more similar. These results indicate that human face recognition is sensitive to statistical information derived from a biologically inspired artificial face recognition system which mimics V1 similarity. The study provides evidence that higher face recognition areas in humans not only preserve, but also possibly compute upon earlier V1 similarity space.