Face recognition with gabor-filter representation
John D. Van Horn, Peter Kalocsai · 1998
It is demonstrated in a set of experiments that the performance of a biologically inspired recognition model that uses the activation of multiscale and multiorientation Gabor-filters as its representation correlates very highly with human performance on various face recognition tasks. It is argued that the preservation of early filter activations even at higher levels in the human visual system is the reason for such a high correlation beside using the appropriate representation type (Gabor-filters) to begin with. All of these results qualify the Gabor-filter system as the currently available best working model of biological face recognition. Another set of experiments demonstrated the differences between face and object recognition in utilizing V1-type early filter activation values. For object recognition strong invariance to these values was found whereas face recognition remained to be heavily dependent on the original filter activations. The result of these experiments gave an underlying reason for the numerous differences between face and object recognition that is so extensively discussed in the literature. Although, much of the V1-type early filter similarity space is probably preserved even at higher face processing areas in the visual system this is not to say that information is not processed there any further. It is discussed how much different aspects of face processing are dependent on the statistics of face images. Numerous statistical techniques are applied to large scale face data such as: univariate analysis of variance, linear discriminant analysis, kernel-based density estimation and difference space estimation to show how statistical information might play a part in face recognition tasks. All the covered statistical techniques demonstrated superior performance on various face recognition tasks compared with the baseline Gabor-filter system which did not utilize any statistical information. In addition, the psychophysical validity of the univariate analysis was also demonstrated via human testing on reconstructed images. A natural extension of this study would be to employ some of the above-mentioned statistical techniques for the purposes of gender, race, age, expression and attractiveness classification of face images.