A Hybrid Metric Estimation/Learning Model for K-NN Classifier

We propose the K-NN adaptive metric learning model with the asymptotic variance estimation for the face recognition. Variance estimation from small training samples is crucial for constructing the K-NN classifiers based on the Mahalanobis distance. The MDL criterion is formulated to obtain asymptotically optimal variances of the DCT feature distributions. The metric parameters are discriminatively trained by using the synthesis method to assume patterns between the K-NN classes. The Mahalanobis distance functions to define the K-NN classes are derived from hybrid metric parameters estimation/learning model. We present the simulation results using the ORL and UMIST databases.