Research On Unconstrained Face Recognition Based On Dictionary Learning

Face images which are collected from reality are influenced by unconstrained factors such as illumination change,expression change and occlusion which can affect accuracy of face recognition.Face recognition algorithm of structured low-rank dictionary learning based on Fisher discrimination is put forward,which is based on subspace projection,theory of low-rank matrix recovery and Fisher criterion.Then,in view of problems from small size samples of face recognition,a new method of singular values reconstruction based on mirror symmetry to construct virtual samples is put forward.This paper firstly researches subspace projection methods,sparse representation and low-rank representation theory.Secondly,this paper proposes face recognition algorithm of structured low-rank dictionary learning based on Fisher discrimination in view of shortage of subspace projection methods and algorithm of Fisher discrimination dictionary learning.The improved algorithm utilizes subspace projection method to reduce dimensions of samples and adopts the reconstruction error of samples to constrain relationships between samples and dictionary.Meanwhile,the algorithm applies Fisher criterion to coding coefficients to make the coding coefficients possess discrimination and applies low-rank regularization to the dictionary based on theory of low-rank matrix recovery.And structured sparse coefficients are imposed to guarantee the optimal classification for testing samples.Finally,this paper analyzes problem of small size samples.In view of the shortage of existing methods for constructing virtual samples,a new method of singular values reconstruction based on mirror symmetry is put forward which combines mirror symmetry with singular values reconstruction.Mirror symmetry ensures symmetry complementarity compared with original samples which are affected by angle and posture change.By selecting appropriate number of singular values,the main feature of original samples retained.In case of simulating insufficient sample size of face databases,this paper expands training sample size by the proposed method to do experiments.The experimental results show that the proposed dictionary learning algorithm is robust and feasible when face images are influenced by unconstrained factors such as occlusion in the face recognition.The proposed method of constructing virtual samples can effectively improve the face recognition rate when sample size is insufficient.