Enhancement of one sample per person face recognition accuracy by training sets extension

This paper deals with one sample per person face recognition, also called one sample per person problem. We use three standard methods: neural networks - MLP (multi-layer perceptron) and RBF (radial basis function) network, and SVM (support vector machine) method. These methods are tested on FERET face database. We analyze impact of extending training sets by modified images (of original images) to improve the training process and thus the overall recognition accuracy. The best test results on modified images are compared to the results using multiple (2, 3, 4) original samples in the training sets.