Research On Image Recognition Based On Semi-supervised Non-negative Matrix Factorization

Non-negative Matrix Factorization(NMF),Semi-supervised Learning(SL)and Deep Learning(DL)are several typical machine learning methods in low rank model and feature representation.They have been widely applied in feature extraction,computer vision and pattern recognition and so on.Based on the deep research of NMF,SLand DL,this paper proposes two improved non-negative matrix factorization methods,and by combining deep learning,this paper proposed a multi-layer NMF network method.1.A Non-negative Enhanced Discriminant Matrix Factorization Method with Sparse Regularization(NEDMF_SR).The proposed method not only introduces the sparsity of coefficient matrix into NMF,but also combines the within-class and between-class discriminant information of coefficient matrix as the regularized term to enhance the ability of discriminantion and sparse represetation.The method maximizes the between-class discrete penalty term,and at the same time,minimizes the within-class compact encouraging term and sparse constraint term of the low-dimensional representation.Moreover,the update rules and convergence proof of NEDMF_SR are also presented.Finally,by constructing local classifier and global classifier,a framework based on the NEDMF_SR method for occluded image classification is given in detail.Extensive experiments demonstrate that the proposed NEDMF_SR method is robust to image corruptions(real disguise and random block occlusion).2.A Multi-channel Robust Discriminant Non-negative Matrix Factorization with Soft Label Constraint(MRDNMF_SL).The method considers 3-point requirements: 1)the low-dimensional representation of the data should be consistent with the original data space;2)the low-dimensional representation should be discriminative,in order to better predict the appropriate label;3)the low-dimensional representation should be robust to singular value and noise.For these reasons,the objective function of MRDNMF_SL is constructed.By minimizing the objective function,the data has good data representation ability in the learned subspace.Then,the method is optimized and by setting the soft label matrix,the method can be adapted to semi-supervised and supervised learning.Furthermore,in order to solve the problem that lack of capacity of deep learning to handle individual appearance variations,especially in ornamental occlusion and light factors environment,a Multi-Layer Non-negative Matrix Factorization Net(MLNMF)is proposed,and the element algorithm of MLNMF is the MRDNMF_SL method proposed in this Chapter.And then,through hierarchical feature representation and fusion,enhances the the ability of discriminantion and makes up for the lack of data representation of deep learning.Through the single-layer and multi-layer feature fusion experiments of the MLNMF_DL network,it can be seen that the problem of insufficient ability of deep learning to deal with the destruction of the image is greatly improved.