Research Methods On Multi-Pose Face Recognition Based On Deep Learning

Face recognition has a wide range of applications which provides a convenient solution to the problem of identity verification in modern society.However,due to limitations of face recognition methods and external factors,most face recognition systems can only be used under the circumstances of the user's cooperation.Therefore,face recognition based on unrestricted conditions is a difficult problem in current research.The performance reduction of face recognition caused by face's pose changing is still a huge challenge in particular.To improve face recognition effect in pose changing is of great practical significance for the application of face recognition system in real life.In view of the problem that face pose changing will cause serious interference to face recognition,this thesis aims at improving the performance of multi-pose face recognition.The research focuses on the face synthesis of multi-pose face under single sample and the deep feature extraction of face under multi-sample to explore the effect of different methods on multi-pose face recognition.The main work of this thesis includes the following aspects:(1)In view of small face pose from-15° to+15° changing scenes with single sample,multi pose face recognition method based on face pose correction of hybrid AAM model with improved shape model and extracting LGBPHS feature are adopted.The AAM model is used to locate the feature points of the face and corrected the multi pose face to the front face and the LGBPHS texture feature of the face is extracted to recognize face.The experiment validates the effectiveness of the method in the face of small face pose changing.(2)In view of large face pose from-45° to +45° changing scenes with single sample,multi-pose face recognition based on feature-constrained stacked progressive autoencoder face reconstruction is proposed.This thesis use autoencoder to synthesis the frontal face.On the basis of the face reconstruction gradually,taking into account the problem of local feature information easily lost in face reconstruction in the process of stacked auto encoder,so the hidden layer adds the feature constraint which constrains the same person's face inner features of the multi-pose face equal to each other to the utmost extent.The experiment can verify the effectiveness of the method for the face pose changing within 45° below the lower resolution.(3)Convolutional deep belief networks model is designed for multi-pose face recognition in view of large face pose changing scenes with multi sample.Because traditional manual features have limitations in multi-pose face recognition and deep networks can independently learn the the deep feature of face,a sparse convolution deep belief networks model is designed for multi-pose face recognition.Face recognition experiments are performed on the original multi-pose face images and the reconstructed face images respectively,which can verify that the convolution deep belief network can better improve multi-pose face recognition performance.The experimental results show that deep learning shows good results in the research of multi pose face recognition method,with the powerful learning ability for nonlinear relations.The results of this thesis effectively improve the effect of multi-pose face recognition.