3D Face Reconstruction And Face Recognition

3D face shape can provide an accurate pose and illumination invariant geometry representation,with wide applications in daily life of people,for example,face recog-nition and animation.Existing acquisition of 3D face data depends on professional devices,such as high-cost,high-precision digital scanners and low-cost,low precision depth cameras.Since 3D face data can not be acquired by mobile phone as photos do,this paper presented a 3D face reconstruction method based on facial images,and ap-plied the reconstructed face shapes or depth images captured by some devices to face recognition tasks.The main contents of this paper include the following items:(1)3D face reconstruction with geometry details from a single image;(2)face recognition methods based on 3D face models;(3)face recognition methods based on depth images.3D face reconstruction with geometry details from a single image:Inspired by recent works in face animation from RGB-D or monocular video inputs,we develop a novel method for reconstructing 3D faces from unconstrained 2D images,using a coarse-to-fine optimization strategy.First,a smooth coarse 3D face is generated from a bilinear face model,by aligning the projection of 3D face landmarks with 2D landmarks detected from the input image.Afterwards,using local corrective deformation fields,the coarse 3D face is refined using photometric consistency constraints,resulting in a medium face shape.Finally,a shape-from-shading method is applied on the medium face to recover fine geometric details.Our method outperforms state-of-the-art approaches in terms of accuracy and detail recovery,which is demonstrated on publicly available datasets.Face recognition methods based on 3D face models:Since the pose of 3D face shape is controllable,we propose a face recognition method based on 3D face models.A 3D face model is reconstructed from a facial image,and then used to normalize the pose of face in this image.Finally,we use SoftMax-based loss functions to train CNN models on a facial dataset with normalized pose.After pose normalization,the diversity of facial pose is eliminated and the face recognition features of the same person become more compact,which helps to improve the face recognition rates,and has been proved by experiments on public data set.Face recognition methods based on depth images:Since the input of the above face recognition methods are still facial images,the advantages of 3D face shape are not fully exploited in face recognition tasks.Thus we construct a large-scale RGB-D facial data set by using depth cameras under real application conditions,and train facial recognition models on this data set.We also develop an attribute-aware loss function for CNN-based face recognition.With the help of attributes,this loss function can make CNN to learn more discriminative features.Finally,experimental results show that both depth channel and attribute-aware loss can greatly improve the accuracy of face recognition.