Research On Sparse Theory In Face Recognition Based On Single Training Sample Per Person

Single sample face recognition problem has been a hot and difficult research topic in the field of pattern recognition, artificial intelligence and machine learning. Many face recognition techniques focus on how to improve the accuracy of a recognition system, and largely depend on the. size and representative of training set. However, it seems that most of them ignore the potential problem that may stem from the face database at hand, where there may be only one sample image per person, possibly due to the difficulties of collecting samples or the limitations of storage capability of the systems, etc. Sparse representation (SR) is initially proposed as an extension of traditional signal sampling and representation theory such as Fourier representation and wavelet representation. In the past few years, SR has proven to be an extremely powerful tool for acquiring, representing, and compressing high-dimensional signals and has been successfully applied to solve many practical problems in fields of signal processing, machine learning, and pattern recognition. Recently, compressive sensing (CS) was proposed by D. Donoho, E. Candes, T. Tao et al., and makes SR a breakthrough in real applications.This paper focuses on the single sample problem based sparse representation for face recognition. The goal of this paper is to improve the recognition performance of the single sample face recognition algorithm. The main contributions of this paper are as follows:(1) This paper briefly introduces the background, significance, status and challenges of current face recognition, sparse theory, and single sample face recognition problem.(2) This paper generates the multiple images using Shift, PCA Reconstruction, Mirror-symmetry Transformation and Sampling methods combined with Sparse Representation-based Classification (SRC), to further improve the recognition performance. These proposed methods not only increase redundant samples and enhance the recognition performance but also save the computation and storage costs.(3) Motivated by the recent development of SR, CS and semi-supervised dimensionality reduction (SSDR), a novel SSDR based on l1-graph is presented, namely semi-supervised sparsity discriminant locality preserving projections (SSDLPP). SSDLPP can remarkably improve the recognition performance of discriminant locality preserving projections (DLPP) and locality preserving projections (LPP) even if we have only single training sample and very few extra unlabeled training samples.Experimental results on both two famous face datasets (ORL and Yale) show that the proposed algorithms substantially improve the recognition performance for the single sample face recognition.