Minimum variance method to obtain the best shot in video for face recognition

This paper describes a face recognition algorithm using feature points of face parts, which is classified as a feature-based method. As recognition performance depends on the combination of adopted feature points, we utilize all reliable feature points effectively. From moving video input, well-conditioned face images with a frontal direction and without facial expression are extracted. To select such well-conditioned images, an iteratively minimizing variance method is used with variable input face images. This iteration drastically brings convergence to the minimum variance of 1 for a quarter to an eighth of all data, which means 3.75-7.5 Hz by frequency on average. Also, the maximum interval, which is the worst case, between the two values with minimum deviation is about 0.8 seconds for the tested feature point sample.