Robust Subspace Learning With Occlusion

Subspace learning has attracted considerable interest for years in the studies.High sample dimensionality and short insufficiency of prior knowledge about the valid features are two challenging problems in this field.Face recognition(FR)remains an active topic after comprehensive research over the last two decades not just owe to its great application potential.It offers a good test-bench to reveal how these two key machine learning problems are solvable as large and unambiguous trial databases are available for FR problem.Partially occlusion is a common difficulty arisen in face recognition.So far,several stateof-the-art face recognition techniques based robust subspace learning have been proposed,such as PCA-L1,R1-PCA,L1-LDA and RPCA.Those methods aim improving the robustness of the transitional linear subspace learning algorithms.In this paper,we consider the robust face recognition problem under occlusions.We first analyze the theoretical basis of exist methods,then find the how the robustness to noise is enhanced.finally,we proposed a novel robust model to extract features in face recognition.Moreover we extend our model to classification and propose a sparse representation based classifier.The main contribute of the paper as follow:(1)A principal component analysis method based on block-norm minimization(BlockPCA)is proposed to exploit improve robustness to outliers.Exist methods treat image by its transferred vector form,it leads to the loss of latent information carried by images and loses sight of the spatial structural details.To exploit these two kinds of information and improve robustness to outliers,we propose Block-PCA which employs block-norm to measure the distance between an image and its reconstruction.Block-norm imposes L2-norm constrain on a local group of pixel blocks and uses L1-norm constrain among different groups.In the case where parts of an image are corrupted,Block-PCA can effectively depress the effect of corrupted blocks and make full use of the rest.In addition,we propose an alternative iterative algorithm to solve the Block-PCA model.Performance is evaluated on several datasets and the results are compared with those of other PCA-based methods.(2)A matrix regression-based classification(MRC)is proposed.Sparse representation based classifier are popular in face recognition.We consider the classifier via matrix regression with L21-norm.Specifically,we propose an inner-class sparse representation classification approach in which images are expressed as matrices instead of vectors and adopted to encode more discriminative information than other regression-based methods.In the regression step,a L21-norm based matrix regression model is proposed,which can efficiently depress the effect of occlusion in probe images.In addition,we argue that the corrupted pixels in probe image should not be considered in decision step.Thus,a self-adaptive threshold is employed to dynamically eliminate the corrupted rows in probe image before making decision.Accordingly,an efficient algorithm is derived to optimize the proposed objective function.Further,we replace the L21 norm in MRC by Block norm and propose MRC-Block.Comprehensive experiments on representative datasets demonstrate that MRC-Block is a robust discriminative classifier which significantly improves the performance and efficiency compared with the state-of-the-art methods in dealing with face image with occlusion.