An adaptive metric learning procedure for reconfigurable facial signature authentication

We present an adaptive metric learning procedure with improved generalization of missing training data for facial signature recognition for use in a smart card system. The conventional learning models suffer from degraded recognition rate due to poor estimation of the margin of a decision boundary. Our model employs an image synthesis method to represent missing patterns of unknown classes by using a mixture distribution. The margin of a decision boundary is dynamically adjusted to input patterns obtained from synthesized images with a time-varying mixing ratio. The metric parameters of mixture distributions have been derived from minimization of the negative log-likelihood probability function. The present method effectively reduces the margin of a class with an improved recognition rate from 81.3% to 100%. Furthermore, we examine the margin structure and select the minimum number of support vectors to represent mixture distributions by using the generalized portrait (GP) method