Comparing GMM and Parzen in Automatic Signature Verification- A step backward or Farward?
File is not available to download!
Source: Proceedings of XVIII Congresso Brasileiro de Automatica (CBA’10), September 2010
Authors: J. Montalvao, N. Houmani, B. Dorizzi
Abstract: The use of Gaussian Mixture Models (GMM), adapted through the Expectation Minimization algorithm, is quite
widespread in automatic verification (Biometric) tasks. Its choice is, at a first glance, founded on the good qualities of GMM
models when aimed at approximating Probability Density Functions (PDF) of random variables. But biometric models for verification
are frequently adapted from small sample sets of biometric signals, since with real applications subjects are not willing to
accept long enrollment sessions. This well known constraint rises a problem of balance between model complexity and sample
size. From this perspective, we show, through simple online signature verification experiments, that constrained GMM with fewer
degrees of freedom, compared to GMM with full covariance matrices, provide better performances. Moreover, pushing this argument
even further, we also show that a Parzen model (seen here as a over-constrained GMM) is even better than most GMM,
in terms of Equal Error Ratio (EER).