Research On Feature Optimization Methods For Face Recognition

With the rapid development of information technology,biometric identification technology is being widely applied to the fields of finance,security and other fields which has been widely accepted by the community.Compared with other biometric identification technologies,face recognition technology has the friendly,intuitive,and reliable characteristics,so the face recognition technology has become a more representative technology in biometric identification technology.In recent years,face recognition technology has made great progress and many classical methods have emerged.However,since the face image is easy to be affected by the illumination,expression and other factors,there are still many challenges in the application of face recognition technology.It involves image preprocessing,dimensionality reduction and classification in face recognition.How to effectively use the relationship between the feature and classifier,thus enhancing the discrimination and improving the concision of features has become a hot issue in current research.This thesis focused on the perspective of Representation based Classifier,by means of feature enhancement and feature extraction,the interaction and restriction between feature and classifier are realized.In addition,based on the feature self representation model,unsupervised feature selection is implemented by linear representation of features and inner product constraints.The main contents of this thesis are summarized as follows:1.A new method of filter learning is proposed,named Supervised Filter Learning for Representation Based Face Recognition.This method can learn the more discriminant local features that minimize the difference of the features between images of the same person and maximize that between images from different people.The characteristics of this method are as follows:(1)Our filter is designed for Local Binary Pattern which can maximize the discrimination ability of LBP features.(2)The linear regression method is used to characterize the intra class and inter class representation errors,and the filter is obtained under the constraint of linear discriminant analysis,so that the filtered features can get better recognition results with representation based classifier.(3)Different from the traditional filters,such as Mean Filter,our filter is learned in a data-driven way.(4)In order to verify the effectiveness of the proposed SFL method,we performed a large number of experiments on both single and multimodal face databases.The experimental results show that our method can effectively improve the discriminative ability of LBP features,and can get better recognition results with representation based classifier.2.A new feature extraction method based on dictionary learning is proposed,named Jointly Learning Discriminative Dictionary and Projection for sparse representation based classification method.This method can get both discriminative dictionary and discriminative features with lower dimensions.The characteristics of this method are as follows:(1)The discrimination of the dictionary is achieved by imposing Fisher criterion on the coding coefficients.Meanwhile,by imposing constraints on the low-dimensional samples,we can get the discriminative ability of the projection matrix.(2)The joint learning makes the learned projection and dictionary better fit with each other so that a more effective face classification can be obtained.(3)An iterative algorithm is accordingly designed with proved convergence to efficiently solve the constrained optimization problem.(4)Experiments on several benchmark face datasets and video databases demonstrate that the proposed JLDDP model is effective for face recognition.Experimental results show that the proposed method can effectively improve the concision and discriminative ability of features.Better recognition performance can be obtained even if the number of training samples is small.3.A new method of unsupervised feature selection is proposed,named Inner product Regularized Nonnegative Self Representation for Unsupervised Feature Selection.This method can remove irrelevant and redundant features,so it can obtain the feature subset with high sparsity and low redundancy in the feature selection.The characteristics of this method are as follows:(1)The self-representation model is used to describe the attributes of the features,thus getting the weight matrix.(2)To ensure the high sparsity and low redundancy simultaneously,we introduce a novel regularization term into the objective function with the nonnegative constraints imposed.(3)The nonnegative constrain of learned feature weights is imposed to strengthen its physical significance.(4)An iterative algorithm is accordingly designed with proved convergence to efficiently solve the constrained optimization problem.Experimental results show that the proposed method can not only improve the concision of feature,but also achieve better classification and clustering results.In summary,the feature optimization problem is studied extensively and deeply in face recognition model.This thesis proposes three feature optimization methods for face recognition which focuses on how to enhance the discriminative of local feature(LBP),how to improve the feature subset of the conciseness and effectiveness,and how to improve the conciseness and discriminative of features by learning projection matrix.It can be seen from the experimental results that the proposed methods have a certain role in promoting the research of representation based face recognition with a good application prospect.