Research On Robust Face Positive Method

Face Recognition has been one of the most active research topics in computer vision and pattern recognition.Although face recognition with frontal view achieves a high recognition rate,the performance of face recognition across pose is poor.To address this problem,we mainly focus on the face frontalization problem based on statistical analysis,three-dimension modeling,feature representation and regression.This paper studies and summarizes the state-of-the art face frontalization methods.Thus,we further improves existing algorithms to promote the effectiveness and robustness of face frontalization method.The main work and research results are as follows:(1)For 2D plane based methods,this paper proposes orthogonal Procrustes regression based on Shatten-p norm.Compared with Frobenius norm,Schatten-p norm could preserve structural information of face images well.We also use l1 norm and l2 norm to constrain the coefficient and provide the corresponding optimization algorithms.(2)This paper proposes a robust face frontaization method based on 3D model.Firstly,we employ SDM(Supervised Descent Method)to detect the feature points.At the same time,given a frontal 3D model and manually localize the feature points.The projection matrix can be estimated according to the two groups of feature points.Subsequently,the face pose can be corrected to the frontal view preliminarily by using the projection matrix.We then fills the invisible region via the local symmetry of face.Finally,we combine the Poisson Image Editing and local symmetry of face to remove the occlusions.