Face Frontalization Algorithm

Face frontalization is the technology that recovers the shape,texture and other information to synthesize the frontal face based on one or more profiles.Face frontalization which is an important part of face recognition systrm,can make the face recognition rate on the non-frontal face condition increase significantly.And the face frontalization is inseparable from the support of face alignment which is the basis of the face frontalization,most face frontalization algorithm requires strict face alignment.To synthesize the frontal face from one profile,we first study the face alignment algorithm,and then on the basis of strict alignment of human face,study the three-dimensional face reconstruction algorithm based on 3D deformation model(3DMM),and finally complete synthesis of frontal face.The main work of this paper is as follows:Firstly,a binocular localization algorithm based on Softcascade and Naive Bayesian classifier with anti-black-box glasses interference and high precision is designed and implemented.And then use the Softcascade algorithm to train the binocular classifier to get the candidate points of the eyes.On this basis,the HOG features of the left,right eye and binocular areas are extracted and their feature spaces are constructed by feature screening,and then the naive Bayesian classifier Obtain the precise positioning of eyes.Then,based on the binocular localization algorithm and close study of the application of supervised descent method(SDM)in face alignment,an improved algorithm is proposed:based on binocular pre-positioning and SDM face alignment algorithm.By adding the eyes pretargeting processing in the process of face alignment initialization,it can effectively avoid the feature points convergence to the local optimal solution in the optimization process and improve the accuracy of face alignment.Experiments show that the improved algorithm proposed in this paper can effectively improve the positioning accuracy of feature points and perform well in various attitude angles(guaranteed binocular view)and common expression conditions.Finally,using 3DMM-based 3D face reconstruction to synthesize the frontal face.Firstly,the SDM optimization method is applied to the training of face posture estimation regression model,to propose the SDM-based face pose estimation algorithm.Then,the matching of 3DMM and profiles is realized by using this method and the face alignment algorithm to reconstruct the 3D face,and finally the frontal face synthesized by using of posture rotation and image repair method based on face symmetry.The experiments show that the similarity between the synthesized frontal face and the original frontal face is high,and the recognition rate of face recognition can be improved effectively.