Modeling Marginal Distributions of Gabor Coefficients: Application to Biometric Template Reduction

Gabor filters have demonstrated their effectiveness in automatic face recognition. However, one drawback of Gabor-based face representations is the huge amount of data that must be stored. One way to reduce space is to quantize Gabor coefficients using an accurate statistical model which should reflect the behavior of the data. Statistical image analysis has revealed one interesting property: the non-Gaussianity of marginal statistics when observed in a transformed domain (like Discrete Cosine Transform, wavelet decomposition, etc.). Two models that have been used to characterize this non-normal behavior are the Generalized Gaussian (GG) and the Bessel K Form densities. This paper provides an empirical comparison of both statistical models in the specific scenario of modeling Gabor coefficients extracted from face images. Moreover, an application for biometric template reduction is presented: based on the underlying statistics, compression is first achieved via Lloyd-Max algorithm. Afterwards, only the best nodes of the grid are preserved using a simple feature selection strategy. Templates are reduced to less than 2 Kbytes with drastical improvements in performance, as demonstrated on the XM2VTS database.