|
Ph.D. Defense Friday, June 1, 2007 Room 216 ENG II 8:30AM Title: Performance Prediction for Biometrics Recognition Systems Abstract: Performance prediction is a fundamental problem for a biometrics recognition system. In this thesis, we address the problem of biometrics performance prediction from a small gallery, learning the optimal small gallery size, and the performance prediction for the multisensor fusion. We present a binomial model to predict a large population performance based on a small gallery. In order to model the distortion happened in large populations, we use a two-dimensional model that combines a hypergeometric probability distribution model with a binomial model to predict a large population performance from a small gallery. By an iterative learning process, we find the optimal size of a small gallery. We use the Chernoff inequality and the Chebychev inequality to determine the small gallery size in theory. We present two theoretical approaches to predict the sensor fusion performance that allow us to select the optimal sensor combination. Both of our prediction approaches are based on the score level fusion. We assume that the match score and the non-match score distributions of the biometrics are mixture of Gaussians. In the first approach, we novelly apply the Neyman-Pearson theory to find the optimal fusion combination instead of doing the Brute-Force experiments. In the second approach, we use a transformation to map a ROC curve to a 2D straight line and derive a metric to evaluate the sensor fusion performance. |
| < Prev | Next > |
|---|