Research On Pose-Robust Face Recognition

Face recognition technology has a very broad application prospect,which is mainly applied to such areas as identity authentication,intelligent video surveillance and human-computer interaction.However,pose variation remains to be one big factor influencing the accuracy of face recognition,so this thesis aims to conduct research on this problem and increase the accuracy of face recognition rate through face frontalization.This thesis studied methods for face frontalization of face image at a deflection angle varying from ±5°to ±25°and PCANet-2(4 Block + Global)face recognition,proving that the methods improved in this thesis have increased the accuracy of face recognition rate on the original basis.The main research contents are as follows:(1)Pose is of great importance to face reorganization.When three-dimensional model is used to frontalize two-dimensional face images,the frontaliztion effect will be far from perfect if depth coordinates are not estimated precisely.As for this problem,this thesis researched registration methods for Candide-3 three-dimensional face models and two-dimensional face images,and depth coordinate estimation methods for the 113 vertices of Candide-3 model in the course of registration.(2)After frontalizing the profile face images at a deflection angle over 5° through Candide-3 three-dimensional face model,the original deflected side will be partially distorted.As for this problems,this thesis researched the detection of distorted surfaces and the restoration of distorted surfaces through constrained texture synthesis after frontalizing the profile face images at a deflection angle from ±5°to ±25°through Candide-3 three-dimensional face model.(3)Based on PCANet-2 face recognition and PCANet-2(4 Block)face recognition put forward by Tsung-Han Chan and Dongliang Liu respectively,this thesis studied PCANet-2(4 Block + Global)face recognition.The specific method is as follows: divide face images into five areas including left eye,right eye,nose,mouth and the whole face,use images from the five areas to train the first-stage and second-stage PCA filters respectively,divide each face image to be detected into five areas,then use corresponding PCA filters to extract features of each areas,and finally recognize the whole face image,whose features are represented by the cascading of the five areas.(4)The experiment shows that when the false alarm rate ranges from 1% to 8%,the PCANet-2(4 Block + Global)face recognition rate is higher than PCANet-2 face recognition rate put forward by Tsung-Han Chan,and when the false alarm rate ranges from 0.5% to 8%,its face recognition rate is higher than PCANet-2(4 Block)face recognition rate put forward by Dongliang Liu.What's more,face recognition rate indeed increases after frontalizing profile face images at a deflection angle from ±5° to ±25°.