A New Quadratic Classifier Applied to Biometric Recognition

Abstract

In biometric recognition applications, the number of training examples per class is limited and consequently the conventional quadratic classifier either performs poorly or cannot be calculated. Other non-conventional quadratic classifiers have been used in limited sample and high dimensional classification problems. In this paper, a new quadratic classifier called Maximum Entropy Covariance Selection (MECS) is presented. This classifier combines the sample group covariance matrices and the pooled covariance matrix under the principle of maximum entropy. This approach is a direct method that not only deals with the singularity and instability of the maximum likelihood covariance estimator, but also does not require an optimisation procedure. In order to evaluate the MECS effectiveness, experiments on face and fingerprint recognition were carried out and compared with other similar classifiers, including the Reguralized Discriminant Analysis (RDA), the Leave-One-Out Covariance estimator (LOOC) and the Simplified Quadratic Discriminant Function (SQDF). In both applications, using the publicly released databases FERET and NIST-4, the MECS classifier achieved the lowest classification error.