Dual-View Normalization for Face Recognition

Face normalization refers to a family of approaches that rotate a non-frontal face to the frontal pose for better handling of face recognition. While a great majority of face normalization methods focus on frontal pose only, we proposed a framework for dual-view normalization that generates a frontal pose and an additional yaw-45° pose to an input face of an arbitrary pose. The proposed Dual-View Normalization (DVN) framework is designed to learn the transformation from a source set to two normal sets. The source set contains faces collected in the wild and covers a wide scope of variables. One normal set contains face images taken under controlled conditions and all faces are in frontal pose and balanced in illumination. The other normal set contains faces also taken under controlled conditions and balanced in illumination, but in 45° pose. The DVN framework is composed of one face encoder, two layers of generators, and two sets of discriminators. The encoder is made of a state-of-the-art face recognition network, which is not updated during training, and it acts as a facial feature extractor. The Layer-1 generators are trained on both the source and normal sets, aiming at learning the transformation from the source set to both normal sets. The trained generators can transform an arbitrary face into a pair of normalized faces, one in frontal pose and the other in 45° pose. The Layer-2 generators are trained to enhance the identity preservation of the faces made by the Layer-1 generators by minimizing the cross-pose identity loss. The discriminators are trained to ensure the photo-realistic quality of the dual-view normalized face images generated by the generators. The loss functions employed in the generators and the discriminators are designed to achieve satisfactory dual-view normalization outcomes and identity preservation. We verify the DVN framework on benchmark databases and compare with other state-of-the-art approaches for tackling face recognition.