Pose Correction and Subject-Specific Features for Face Authentication

In this paper, we present a face authentication system that can be broken down in three stages. Prior to feature extraction, a pose correction step is applied, so that frontal face images are synthesized through the use of flexible shape models and thin plate splines warping. Once the pre-processing is completed, a set of subject-specific key points is located in face images by means of a two-layer strategy. In the first level, locations are chosen from lines that depict facial structure, and Gabor features (jets) are extracted at these positions. In the second step, we evaluate these jets as individual classifiers over a training dataset, so that only the subset of points whose jets are good at discriminating between clients and imposters are preserved