Face Recognition Algorithms Based On Incremental Non-negative Matrix Factorization

With the continuous progress in social informatization,the face recognition technology has already surpassed the level of theoretical research and has been used in a wide range of fields.From the field of criminal identification in public security system to the field of identity recognition in automatic attendance system,the demand for face recognition application has increased steadily,and face recognition has become one of the most promising biometric recognition technologies.Non-negative Matrix Factorization(NMF)is a classical subspace dimensionality reduction algorithm,which is widely used in the field of face recognition due to its good explanatory ability.However,the traditional NMF based face recognition algorithm usually adopts batch learning,and the data learning speed will lag behind the update speed when the face data scale up incrementally.Therefore,in order to meet the requirements of online face recognition learning,it is particularly important to use the incremental learning technique to continuously learn new knowledge from the new training samples and instantly update the original model parameters.With the aim to apply incremental learning for small sample database,this thesis proposed several improved algorithm models for online learning and obtained the following results:1.the research background,algorithm model and optimization solution method of NMF was introduced.Several mainstream NMF-based incremental learning models was further introduced,compared and analyzed,including Incremental the Non-negative Matrix Factorization algorithm(INMF),the Incremental Graph Regulated Non-negative Matrix Factorization algorithm(IGNMF)and the Incremental Orthogonal Projection Non-negative Matrix Factorization algorithm(IOPNMF),etc.2.In the Traditional incremental NMF,factor matrix lacks sparsity,and the image's local expression ability is insufficient.To solve these problems,an Incremental Non-negative Matrix Factorization with Sparseness Constraints(INMFSC)was proposed.In this method,sparse constraints were applied to the objective functions of the traditional INMF algorithm,and sparse weight coefficients were imposed to the basis matrix and coefficient matrix after factorization.Experimental results on both ORL and CBCL face databases showed that the proposed method improved the sparsity of factor matrix,greatly enhanced optimization solution speed,and showed better local expression ability on face base image compared with the traditional algorithm.3.It is difficult for the traditional unsupervised INMF algorithm to achieve high clustering accuracy and abundant normalized mutual information in reduced dimensional subspace.Incremental Constrained Non-negative Matrix Factorization algorithm(ICNMF)was proposed to overcome those problems.Based on semi-supervised learning Constrained Non-negative Matrix Factorization(CNMF),ICNMF implemented semi-supervised incremental learning by using the idea of partitioned matrix.It not only has merits of the traditional INMF,but also can carry out incremental updates for factor matrix according to whether new samples carry tag information or not.Experiments on ORL,Yale and PIE_pose27 face databases proved that the proposed algorithm has higher accuracy and normalized mutual information value than the traditional algorithm in clustering.4.For the problems of low recognition rate and convergence speed of the traditional INMF,an Incremental Discriminant Non-negative Matrix Factorization algorithm(IDNMF)was proposed.The INMF is an unsupervised incremental learning method based on subspace dimensionality reduction technology,in which lack of tag information has some negative effects on the effectiveness of face recognition and algorithm convergence.Our approach utilized label information of training samples,and used coefficient vector means of the same class for initialization.Then the algorithm iterated with constraints on minimizing Euclidean distance of within-class samples,and resulted more discriminative features and less iterations in computation.Experiments on ORL and PIE face databases demonstrated that the proposed algorithm achieved better classification accuracy and converged faster than former batch based non-negative matrix factorization algorithms.5.Based on IDNMF algorithm,an Incremental Non-negative Matrix Factorization based on Fisher Discriminant Analysis algorithm(FINMF)with further optimization and improvement of the utilization of tag information was proposed.To further improve the utility of tag information in INMF,the idea of fisher discriminant analysis was adopted,and a new non-negative matrix factorization incremental learning algorithm with discriminative information and constraints was proposed.Firstly,the prior information of original training samples was used to initialize the incremental coefficient matrix through an indicator matrix.Secondly,the object function of INMF was improved to be a batch-incremental learning algorithm,with the constraints of maximized between-class scatter and minimized within-class scatter.Finally,the factor matrices were calculated by the method of multiplicative iteration.Experiment results on the ORL,Yale B and PIE face databases showed that the proposed algorithm has higher recognition rate and lower time cost compared to other congeneric algorithms.