Research On Nonnegative Matrix Factorization Based Occluded Face Recognition

There are many methods for feature extraction and data dimensionality reduction.Among them,Non-negative Matrix Factorization(NMF)has attracted extensive attention and been applied in many fields due to its characteristic that the whole is composed of parts.Face recognition is a good example of NMF application.At present,face recognition is facing many challenges in real life applications,such as poor resolution,light intensity is not fixed,human posture is not fixed,random object occlusion and so on.Among them,occluded face recognition is difficult,did not cause enough attentions.In order to solve this problem,this paper constructs two innovative NMF algorithms for face recognition with occlusion after in-depth study of NMF.(1)Robust Structure Preservation Nonnegative Matrix Factorization(RSPNMF)is constructed.In this method,the Correntropy-induced Metric(CMI)is firstly used to measure factorization error,giving large weight to the elements with small reconstruction error,and small weight to the elements with large error caused by damage or occlusion,thus suppressing the influence of occlusion and noise and improving the expression ability of features.Next,the structure preservation constraint is constructed so that each sample can be reconstructed from its neighbors,thus better capturing local information in the data space.Experimental results show that RSPNMF is a very effective method.(2)Adaptive Local Learning Nonnegative Matrix Factorization with Max-Margin(ALLNMFMM)is proposed.Firstly,a smooth matrix is embedded in NMF to enhance the sparsity of the learned representation.Then,adaptive nearest neighbor learning is used to learn the intra-class adaptive nearest neighbors and inter-class adaptive nearest neighbors of each sample.Then,the max-margin constraint is constructed so that a pair of samples with longer distance in the same class can minimize the distance in the feature space to achieve intra-class compactness,and a pair of samples with close distance between different classes can maximize the distance in the feature space to achieve inter-class separation.Furthermore,large punishment is given to samples with a longer intra-class distance and samples with a closer intre-class distance,so as to maximize the boundary distances of samples of different classes in the feature space.Finally,the effectiveness of ALLNMFMM method is proved by experiments.(3)A Joint Nonnegative Matrix Factorization framework(JNMF)for occluded face recognition is also constructed.Firstly,samples are input into RSPNMF to reduce the dimension,remove the redundancy and extract the discriminative features of the occluded face,and realize the first decomposition.Then the extracted face features are input into ALLNMFMM for secondary decomposition,which makes the samples with the same label more clustered,the samples with different labels more scattered,and mining more discriminative features.The effectiveness of JNMF has been proved by a large number of experiments.