Person classification upon face image based on simile classifiers

A new approach to person classification upon face images is presented. It consists of two steps. First, similarity distances between a test face and faces from a known sample are computed. For this task, special simile classifiers are trained. Similarity distances are estimated independently for each face fragment, e. g., nose, mouth, left or right eye, etc. Pixel colors, brightness, gradient norm, and orientation are used as features both for each pixel and for whole fragment in the form of a histogram or Gaussian distribution parameters. Person classification is performed based on these similarity distances. For classification, support vector machine with radial basis function kernel is used. The proposed algorithm was tested on a gender classification task, using Labeled Faces in the Wild and Public Figures Face databases. The algorithm achieved 92.96% accuracy.