Research On Face Recognition With Pose Variation

In recent years,deep learning has been successfully applied in face recognition and push face recognition application.In practical application,pose variation is one of the important factors that affect the performance of face recognition.In this thesis,the problem of pose variation and single sample in face recognition are discussed.Each subject has a single gallery image with front view and normal lighting.The proposed algorithm works to faces with limited pose variation.The primary work is as follows:We proposed a probability weighted head pose estimation method based on convolution neural network.To address the problem of limited head poses in the training,we propose a head pose estimation scheme to combining multiple label classification and probability weight.The experimental results show that the proposed algorithm could get the 1 degree errors for the pose classifications in the training set,and the 5 degree errors for the untrained pose classification.It is useable in practice.We propose a face recognition scheme using the simulated face image with pose variation,which is generated based on 3D face model.Firstly,the 3DMM method is used to reconstruct the 2D human face into a three-dimensional model.After using the proposed head pose estimation algorithm to get the pose of the face.Then the 3D model is rotated into the estimated pose.Finally,face recognition is implemented by combining ASIFT with SSIM.The experimental results show that the proposed scheme is effective to face images with only left-right rotation.