Making 2D face recognition more robust using AAMs for pose compensation

Abstract

The problem of pose in 2D face recognition is widely acknowledged. Commercial systems are limited to near frontal face images and cannot deal with pose deviations larger than 15 degrees from the frontal view. This is a problem, when using face recognition for surveillance applications in which people can move freely. We suggest a preprocessing step to warp faces from a non frontal pose to a near frontal pose. We use view-based active appearance models to fit to a novel face image under a random pose. The model parameters are adjusted to correct for the pose and used to reconstruct the face under a novel pose. This preprocessing makes face recognition more robust with respect to variations in the pose. An improvement in the identification rate of 60% (from 15% to 75%) is obtained for faces under a pose of 45 degrees