A class density approximation neural network for improving the generalization of Fisherface

Fisherface is a popular dimensionality reduction technique in face recognition. The obtained discriminative subspace, however, may not generalize well to unseen classes, as in the case of enrollment of new identities. In this paper, a class density approximation network based on SOM2 is introduced to derive “prototype classes” for training, which improves the generalization of Fisherface and reduces its performance variance resulting from random selections of training classes. We also propose a splitting method to estimate the number of prototype classes needed in approximation. Experiments on synthesized data and real face databases validate the effectiveness of our method.