Single-sample Face Recognition Problem Based On Intra-class Differences In A Variation Model

Face recognition has advantages in its naturalness and invisibility, however, therecognition performance will be decrease seriously when it refers to the non-ideal imagequality, such as facial instability, expression variations, illumination changes andocclusions etc. Besides, it will be also affect the recognition rates when the face database istoo large or too small. Currently, Sparse Representation-Based Classification (SRC) hasbecome one of the heat topics of the face recognition area. The SRC theory believes that allthe training images are regards as a sparse representation dictionary, and the query facialimage can be represented as a sparse linear combination of the training imagescorresponding to the same person. However, SRC puts forward higher requirement for thesample quantity and it is not robust to facial variations such as illumination, disguise.Single-sample face recognition problem is one of the most challenging problems in facerecognition. More specifically, if there is only a single training face image per person, it isgenerally inaccurate to estimate the intra-personal and inter-personal variations.To address these issues, we challenge the single-sample face recognition problem withintra-class differences of variation in a facial image model based on random projection andsparse representation. In this paper, we separate intra-class difference from facial variationmodel by low-rank optimization and sparse constraint methods. Furthermore, randomprojection matrix project high dimensional facial image onto random subspace. Finally, wepropose a novel facial random noise dictionary learning method that is invariant to differentfaces. The experiment results on the AR, Yale B, Extended Yale B, MIT and FEI facedatabases validate that our method leads to substantial improvements, particularly insingle-sample face recognition problems.