Novel Face Recognition Technology Based On Robust Subspace

Face recognition technology is one of the important topics in the field of biometric identification and artificial intelligence, which has been widely used in many fields, such as finance, public security, social welfare, e-commerce, security and defense and so on. Feature extraction plays the most important role throughout the face recognition process. An effective extraction method can extract the representative features from the face data. And with this, we can classify the data to the corresponding mode which not only simplifies the separator design, but also enhances the accuracy of face recognition. Subspace learning methods which are classical methods in face recognition are catching more and more attention and application because of the advantage in description, recognition accuracy and computation cost. As a subspace segmentation method, low-rank representation can discover the internal structure in face data which gets a lot of attention and promotion. It aims at finding a block-diagonal matrix, in which each block is associated with a sample category. LRR is more robust to noise by introducing an error term, so that it can handle the data in complex case well. In recent years, some methods combined with low-rank representation and subspace learning have been proposed and achieved promising performance. They take the advantage of both LRR and subspace learning which makes them more robust and gain appealing results. In this paper, the study will include the following aspects:(1) Three important subspace learning methods, Principal Component Analysis(PCA), Linear Discriminant Analysis(LDA) and Locality Preserving Projections(LPP), are introduced and analyzed comprehensively. They are linear methods.(2) The low-rank representation(LRR) is discussed after the analysis about classical subspace learning methods. LRR is a Subspace Segmentation method. Unlike Sparse Representation(SR) which seeks the most sparse representation of data from single subspace, LRR seeks the lowest rank representation of data from multiple subspaces. The experimental results show that LRR is more robust to noise.(3) A typical method is analyzed, which combine subspace learning with low-rank representation. The experimental results on face datasets demonstrate that this method can dig out the global structure in face data better and can handle the high-dimensional condition well. And the recognition accuracy has been improved significantly.(4) Based on the analysis about the method combined subspace learning with low-rank representation, a novel metric learning method based on low-rank representation is proposed. Two scenarios of experiments,(e.g. face verification and face identification) are conducted to estimate our algorithm. Experimental results on challenging face datasets reveal the stability and effectiveness of our method.