Research On Local Method For Face Alignment

With the decline in the price of hardware devices and the development of image analysis and processing technologies,face recognition began to be extended from laboratory research to commercial use.As an important part of the face recognition system,face registration connects the two key steps of face detection and feature extraction and matching,which has an important impact on the performance of face recognition.If the face registration accuracy is high,then face recognition performance can be improved,and vice versa.Therefore,the accuracy of face registration plays a decisive role in the face recognition system.Firstly,a face model is established and the model is used as the reference face when registering.Then,the affine transformation model is used to find the correspondence between the reference face and the instance face feature points.Then the Nelder-Mead simplex method is used to search the optimal parameters of the face model for rigid transformation,and a cost function is established to constrain the optimization process.A local face registration model using rigid transformation is constructed.The experimental results show that the registration method can achieve a rough positioning effect on the feature points in the inner region of the face.Aiming at the problem that the method has poor positioning effect on the contour points of the human face,considering the combination of global rigid transformation and non-rigid warp,the registration performance of individual faces in a group is improved by joint registration of a group of faces.Firstly,the supervised descent method is used to determine the initial position of the face calibration point in the group,and then the RANSAC algorithm is used to solve the optimal homography transformation matrix between the face feature points in the group,which is determined according to the optimal homography matrix.Corresponding relationships between facial feature points are determined according to an optimal homography matrix.Then,the feature points in the reference face model of the non-rigid ICP algorithm are non-rigidly warped,so that each feature point in the reference face moves to the true position of the corresponding feature point in the instance face.Finally,a local face registration method using groupwise registration is established.Performing experimental verification analysis on the commonly used face registration database 300 W,compared with the registration result of the supervised descent method,in the case of a positive face,the group registration method can improve the average positioning accuracy of the instance face feature points in the group.The registration method using the groupwise registration method can improve the feature point location accuracy of the frontal face compared with the SDM algorithm.The obtained registration result can lay a foundation for the subsequent high-level visual tasks related to the face,which has important research significance.