Large Pose Face Alignment Based On FDVA Model And Its Application

As an important part of face recognition system,the accuracy of face alignment algorithm will directly affect the final result of face recognition.In unconstrained scenes,face images are affected by various factors,such as illumination,posture,expression and so on,which leads to many problems and bottlenecks in face alignment.At present,most face alignment algorithms are designed for small to medium pose(face deflection angle is less than 45°).When part of the face is self-occluded due to large face deflection,the original face alignment algorithm can't accurately locate the key points of the face.Therefore,how to improve the accuracy of face alignment in unconstrained scenes,especially in the case of large pose angle deflection,has important research significance.Aiming the problem,this paper makes an in-depth study on improving the accuracy of face alignment algorithm in large pose scene.The main research contents and innovations are as follows:Aiming at the problem that the accuracy of face alignment algorithm is greatly reduced due to face deformation and missing in large pose scene,a multi-channel network architecture model combined with multi-task learning is proposed in this paper.Firstly,the task of face alignment is divided into three different parts,face detection,landmark localization and landmark visibility.Through triple independent optimization,FDVA model is established.Then,the model can accurately locate the facial landmarks,and the output results are hardly affected by the self-occlusion caused by posture.Finally,the experimental results show that the normalized mean error(NME)of the model on AFLW2000-3D dataset is 4.12%,which is better than the mainstream face alignment algorithm.Aiming at the problem of low accuracy of pose correction and low degree of detail restoration of face model in large pose scene,this paper proposes a face reconstruction algorithm based on FDVA face alignment.Firstly,through the facial landmarks output from FDVA model,a facial pose estimation model based on facial landmarks is established.Then,according to the results of facial pose estimation,the pose of the average face model is corrected.At the same time,combined with the convolution neural network as a regressor,the mapping relationship between the human face alignment results and the face model is iteratively optimized.After that,the face reconstruction will be realized.Face reconstruction is completed with high accuracy,and the accuracy is hardly reduced by the distortion caused by face posture or expression.Finally,a face model with high degree of restoration and strong fidelity is obtained.The experimental results show that the mean average error(MAE)of the model on aflw2000-3d data set is 4.26%,which is better than the mainstream face reconstruction algorithm.