Nonnegative Matrix Factorization Method With New Characteristic For Occluded Face Recognition

As an effective algorithm for dimensionality reduction and feature extraction,Non-negative Matrix Factorization(NMF)has been widely applied in many fields.Deep learning,especially the deep learning network with attention mechanism,has become the most popular intelligent learning algorithm at present.Base on the in-depth study of NMF,attention mechanism and deep learning,a face attribute prediction method based on attention mechanism is proposed in this thesis.And furthermore,an improved NMF method based on attribute embedding is proposed.(1)Face Attribute Prediction Network with Symmetric Attention Mechanism(FAP-SA)is proposed.The symmetry constraint is introduced into the basic attention mechanism,and the acquired attention modules are integrated into the traditional convolutional neural network.Then,a unified framework for predicting multiple face attributes is constructed.Moreover,the construction of objective function and training process for FAP-SA are also described.Finally,FAP-SA is fine-tuning on the standard face attribute prediction database,and the experiments of feature visualization and prediction for face attributes are carried out.The experimental results show that the proposed FAP-SA network has satisfactory performance in face attribute prediction tasks.(2)Sparse Attributes Embedding Nonnegative Matrix Factorization is proposed.Firstly,the probabilities of face attributes are obtained from the proposed FAP-SA network,and the attribute probability matrix is constructed to achieve high-level semantic information embedded for face images in NMF.Then,sparsity constraint is constructed to obtain the objective function of SANMF,and corresponding iterative update rules and convergence proof are given accordingly.Moreover,the framework and process of applying SANMF for face recognition under occlusion are given.Furthermore,SANMF is used as a unit algorithm,and a Deep SANMF network is proposed to effectively acquire hierarchical features.Finally,experiment results prove the effectiveness of the SANMF method and the Deep SANMF network.(3)A two-level feature fusion network is also proposed,which is utilized to build Deep Attention combined NMF face recognition framework(Deep Att-NMF).In view of theredundancy of face attribute features obtained by FAP-SA network,sparsity constraint is introduced to achieve the first-level feature fusion.After that,the fused features are combined with the features of Deep SANMF for the second level to achieve a more comprehensive occluded face recognition.The experimental results fully demonstrate the effectiveness of Deep Att-NMF.