Research On Face Recognition Based On Local Fisher Criterion And Subspace Analysis

Face recognition is an activate research field in pattern recognition and computer vision because of its immense application potential and significant academic value.How to extract effective feature from original face images and recognize the objection is a key issue in the procedure of face recognition.Since only small numbers of high-dimensional training samples are available,it's hard to recognize the target.Dimensionality reduction is inevitable in solving computer vision and pattern recognition problems.As a dimensionality reduction and feature extraction technique,linear subspace is a popular research hotspot in face recognition.Sparse representation also follows parsimony principle and it's robust to corruption signals.Preserving the robustness to noise,a special linear subspace learning method based on sparse representation which directly maps the samples to a low-dimensionality space is an easy and effective method.Inspired by linear subspace and sparse representation,this dissertation proposes several novel feature extraction algorithms.The main works and distributions of the dissertation are as follows:(1)The local structure between face samples has significant influence on the performance of face recognition algorithms.To overcome the problem of locality neighbor parameter selection in locality preserving projections and take advantage of samples labels information,fisher locality discriminant embedding is proposed,which can preserve more accurate local neighbor relationship and increase the discriminant ability after reducing the sample dimension.For preserving the intrinsic geometry distribution of the samples data.,the proposed algorithm adaptive learns locality neighbor parameter k through sparse representation which reconstruct each sample,e.g.a face image,using as few training samples as possible.Based on the sparse locality neighbor relationships of the samples,the proposed algorithm constructs intra-class compactness and interclass scatter matrix,which can preserve sparse locality neighbor relationships and improve the discriminant power by class labels information.By solving the constructed fisher locality discriminant embedding optimization problem,data points of the same class maintain their intrinsic neighbor relations,whereas neighboring points of different classes no longer stick to one another.The experimental results on benchmark face images datasets(Yale,AR and Yale B)show that the proposed algorithm has higher recognition accuracy than the existing algorithms in subspace feature extraction and validate the efficiency of proposed algorithm.(2)To solve the problem of extracting subspace directly from original face samples by statistics methods,which is easy to be affect by detrimental contribution features and noise in the original samples,a novel linear subspace learning algorithm based on discriminant dictionary learning for face recognition is proposed.Firstly,to improve the discriminant of dictionary atoms and preserve the local structure of sparse coefficients,a structured dictionary whose dictionary atoms have correspondence to the class label is learned,and local structure of the same class samples is preserving,meanwhile of the difference class samples is preventing in the sparse coding coefficient.Secondly,in order to weaken the influence of noise and detrimental contribution features calculated from the original samples,the reconstructed samples were used to extract discriminant linear subspace through MMC which can conquer the small sample size problem in discriminant feature extraction.The experimental results on benchmark face images datasets(Yale,AR and Yale B)show that the proposed algorithm outperforms the conventional algorithms in subspace feature extraction.