Research On Facial Landmark Localization Algorithms

Face alignment is a method to locate semantic facial landmarks such as eyes,nose,mouth and chin with many discrete coordinate values.It is essential for tasks like face recognition,face tracking,3D face modeling,face animation and other important areas.With the explosive increase in personal and web photos nowadays,a fully automatic,highly efficient and robust face alignment method is in demand.In this paper,to improve accuracy and speed of face alignment,some improved existing algorithms are proposed which are listed as follow:1.In order to overcome the problem that explicit shape regression has low precision in face alignment,an improved explicit shape regression for face algorithm is proposed.Firstly,to get a more accurate initial shape,three-point face shape is used as an initial shape mapping standard to replace face rectangle.Then,pixel block feature is used against illumination variations instead of pixel feature,which improved robustness of algorithm.Finally,instead of average method,the accuracy of algorithm is further improved by multiple hypothesis fusion strategy which merged multiple estimations.Compared with explicit shape regression algorithm,simulation experiment results show that accuracy is improved by 5.56%,3.26%and 1.93% respectively using the algorithm proposed in this paper on LFPW,HELEN and300-W face datasets.2.The cascade regression model has low precision and slow running problem in locating facial landmarks because of dependent shape initialization and high complexity,so a face alignment algorithm via affine-transformation parameters regression and local binary features is proposed.Firstly,face shape is initialized by affine transformation parameter regression,and initial shape of the transformed shape is closer to the ground truth for improving model accuracy and convergence speed.Then,random ferns that follows local learning principle are constructed in local area of each feature point,and then learn to obtain easily and highly sparse binary feature to improve speed of model.Finally,global linear regression is used to compute binary feature and obtain shape increment to complete face alignment.Experimental simulation results show that accuracy is improved by 5.03%,3.17% and 1.49% and consuming time is reduced by 0.08 s,0.02 s,0.09 s respectively in this paper compared with existing algorithms on LFPW,HELEN and AFW datasets.3.Aiming at the problem that low precision when locating facial landmarks due to faces with large view variations,a multi-view face alignment algorithm is proposed.Cascaded Pose Regression(CPR)model which combine random forest local learning principle and global linear regression is used to establish different models under multi-view faces.Multi-view models is used to improve accuracy of face alignment to replace single model.Firstly,CPR model is used to establish different models for multi-view faces.Then,multi-view generative model(MVGM)is used to evaluate head pose for input face.Finally,according to head pose,a corresponding model is selected for input face to achieve high precision of face alignment.Experimental results show that face alignment algorithm in this paper has higher location precision than several existing face alignment algorithms.