Research On Pose-Invariant Face Recognition

Many fields start to promote the intelligent innovation increasingly with the continuous evolution and gradual implementation of artificial intelligence technology.Biometric identification is a kind of technology which is used to identify individuals through physiological or behavioral characteristics.Face information plays an important role in the field of biometric recognition because of its unique matching with individual identity.One of the knotty problems in face recognition is that the accuracy drops a lot in the case of recognizing faces captured in varieties of poses,which directly affects the application range of face recognition technology.Therefore,this thesis will focus on reducing the interference caused by facial posture changes on face recognition.Under the condition that the facial posture changes,this thesis focuses on the multipose face image preprocessing algorithm and the pose robust feature extraction algorithm.Improving the accuracy of face recognition in unconstrained pose is the main purpose of this thesis.The main work is summarized as follows:1.The basic framework of pose independent face recognition is introduced,and the key technologies involved in each part of the framework are compared and analyzed.2.The pose normalization algorithm based on texture filling and 3D face model is proposed in this thesis.Aiming at the problem that the traditional pose normalization algorithm requires a large number of samples of the same individual to train different angles of view in different poses,this thesis introduces the 3D model of the face and establishes the homography transformation relationship of the local texture region of the face image.Face images captured in different poses are reconstructed in frontal situations.Aiming at the problem that the local texture is missing in the face image with large pose after the pose normalization,the registration relationship between the single image and the three-dimensional deformation model is studied,and the Poisson editing algorithm is used to reasonably complete the missing texture.3.A face recognition algorithm based on fusion of multi-source filter bank features is proposed.This method firstly establishes Gabor filter banks and ICA(Independent Component Analysis)filter banks in different sizes,and obtains abundant local feature information through traversal convolution.The features extracted from a single filter bank are redundant and the facial information is not fully characterized.Therefore,two types of features derived from different filter banks are fused by using the theory of canonical correlation analysis.Experimental results show that the features obtained by the fusion method in this thesis are robust to face images with pose changes,which improves the recognition accuracy of face images with different poses.