Research And Implementation Of 3D Face Alignment Method Using Convolutional Neural Network

With the improvement of human awareness of property security and the development of image analysis and processing technology,face recognition technology has been gradually promoted to the business model,used in all walks of life,such as Alipay can use face to pay,face access control system,etc.Face alignment is an important part of face recognition field.It connects the two important steps of face detection and feature extraction,which has a huge impact on the performance of face recognition.When the result of face registration is good,the performance of face recognition is improved effectively.Early face alignment methods consider two-dimensional faces,which usually cannot represent the change of face depth and ignore the depth information of the face,resulting in the problem of shape inconsistency when the face rotates in three-dimensional space.Because the face can usually be rotated in three dimensions,the use of two dimensional alignment method is prone to a large error.The 3D deformation model is introduced to improve the robustness of face rotation.For this reason,this thesis uses a 3D face alignment algorithm.The algorithm uses the end-to-end idea,inputting only one face image to complete the face alignment,using the PRN(Position Map Regression Network)structure to locate the key points and obtain the 3D information.In order to further optimize the parameters of the network model,the same weight ratio of different facxe regions was set,and the loss function was improved.The joint alignment of two-dimensional face images is considered,and the alignment process is divided into two stages: affine transformation and non-rigid distortion.Facial point on the affine transformation phase is divided into five groups,face contour,the left eye,right eye,nose and mouth,respectively for each group of iteration,the phase reference face dynamic update,to face all the key points in nonrigid distortion phase do not distort rigidity,but in different people consider the local differences between the key points on his face.This method is superior to the method which only considers the global affine change and the method which considers the global affine change and non-rigid body distortion.In addition,the algorithm in this thesis makes use of the grouping of key points to avoid the contraction of the alignment results after groupwise face alignment.The performance of the proposed method is verified by experiments.The proposed method takes into account the global and local changes of human face and can provide strong support for face alignment and facial expression analysis.