Research On Face Recognition Based On Non-negative Matrix Factorization

In recent years,face recognition has been widely used in video surveillance,financial payment and other fields,reflecting important business value and application prospects.But how to accurately and effectively conduct face recognition and improve information security has become an important research project.As a typical subspace analysis method,Non-negative Matrix Factorization(NMF)satisfies the application requirements of face recognition.At present,the face recognition algorithm based on NMF method has effectively improved the recognition accuracy.However,the face images acquired in practice are disturbed by noise and obstructions,resulting in a decrease in recognition accuracy.Based on the analysis and summary of the existing NMF methods,this study proposes two improved methods for face recognition.The main work of this study is as follows:1.Pinpointing to the insufficient learning ability of the existing NMF methods for local face information,the study proposes A Local Discriminant Non-negative Matrix Factorization(LDNMF)method.By designing the weight operator,the local coordinate constraint is weighted,which is used as a local constraint term in the objective function.The constraint focuses on those face images with greater differences,making the base vector closer to the original data.At the same time,in order to learn more distinguishing features,the inter-class operator is introduced into the coefficient matrix and used as a discriminatory constraint in the objective function.By introducing these two kinds of constraints,the LDNMF method in this study is more suitable for practical application scenarios.Meanwhile,for the sake of weakening the influence of illumination and occlusion,this study designs a face recognition algorithm by combining Gabor feature map and LDNMF method,and gives detailed implementation steps.The method proposed in this thesis can effectively improve the recognition result2.Concerning the problem of correlation between different perspectives which cannot be well explored for existing multi-view methods,a Multi-view Non-negative Matrix Factorization(Multi-view NMF)method is presented.By designing a multi-angle constraint operator,the difference matrix is calculated for the two different viewing angle coefficient matrices.The sum of norm sum squares of all the difference matrices is introduced as a multi-view constraint term into the objective function,and the constraint focuses on learning the consistency information and complementary information among different perspectives.In order to reduce the interference of similar face images between different categories,the inner products of different class matrices from the same perspective are used as category constraints in the objective function.Through introducing these two kinds of constraints,the multi-view NMF method can learn more multi-view information.This study combines multi-scale concepts and Multi-view NMF methods to propose a multi-scale multi-view face recognition algorithm so as to learn face information at different scales,and gives detailed implementation steps.The experimental results show that compared with other methods of the same kind,the proposed algorithm performs better in learning multi-scale and multi-view information,and also improves the accuracy of human face recognition.