Uncertainty evaluation in face recognition algorithms

The paper proposes a method that takes into account the measurement uncertainty in pattern recognition procedures, where, generally, an input is classified searching the most similar, by means of some quantitative parameters, in a database of reference to the comparing the unknown. The result of the comparison between the measured values and the reference ones is not deterministic because of the uncertainty on both the value sets. As a consequence, the decision (recognition of subject) has a risk level, thus it might be wrong. The proposed approach is focused to give a quantitative assessment of the measurement uncertainty and consequently the risk level in decision-making. The case study refers to the face recognition with the Linear Discriminant Analysis (LDA) approach. The recognition is performed by comparing the values obtained with LDA algorithm on observed images and those obtained applying the same LDA to stored reference images.