This paper proposes a feature selection method that aims to select an optimal feature subset to representing facial image from the point of view of minimizing the total error rate (TER) of the system. In this proposed approach, the genuine user score distribution and the imposter score distribution are modeled based on a Parzen-window density estimation to enable the direct estimation of total error rate (TER) as reflected by the area under the curve of the overlapping region of both distributions. Particle swarm optimization (PSO) is employed to search for feature subsets which are extracted from discrete cosine transform or principal component analysis that gives minimum TER and in the meantime to reduce the dimensionality of the feature set thereby reducing processing time.