Face Pose Correction With Eyeglasses and Occlusions Removal

Abstract

This paper presents how to remove unwanted occlusions such as eyeglasses in face images. By choosing training images carefully, we can derive a set of basis vectors that can eliminate this impediment for processing face images. In order to handle different poses, we apply multi-view active appearance models (MVAAMs) for fitting and then convert non-frontal images into frontal neutral faces. A set of these frontal-neutral faces is chosen for the basis and used for reconstructing other face images with occlusions. In addition, we are able to correct missing features while converting into frontal faces. The corrected faces are used for inputting to a frontal based face recognition system which can handle non-frontal faces efficiently.