Research And Application On Discriminant Model Of Sparse And Low-Rank Representation

With the development of society and the progress of science and technology, face recognition technology is becoming increasingly significant, which shows extensive application prospects. At present, researchers have proposed a variety of new approaches with respect to face recognition. Meanwhile, the studies also obtained a certain achievements. Whereas, several aspects contribute to the difficulty of FR problem including the large variability in variance, illumination, pose, occlusion and even disguise of different subjects. So the face recognition technology has enormous significance of academic research and broad development space. Researchers always pay attention to feature extraction of facial images and the generalization of classifiers for the design of practical face recognition systems. With the rapid development of Sparse Representation (SR) and low-rank representation (LRR), various kinds of face recognition algorithms based on SR or LRR has been proposed, which show strong vitality. Although the testing sample might be corrupted, the training data sets are commonly assumed to be well taken in some desired conditions including reasonable illumination, pose, variations and without occlusion or disguise. When applying existing face recognition methods for practical scenarios, we will need to throw away the corrupted training images, and we might thus encounter small sample size and over-fitting problems. Moreover, overlooking the damaged trained face images will loses the information that value to recognitions. Therefore, when such noise is presented in both training and testing data, how to sufficiently extract the discriminant information from training samples to improve the identifying performances is still a research hotspot.This paper mainly studies the discriminant model based on sparse and low-rank representation. The goal of this paper is to achieve an efficient discriminant model that can improve the performance of face recognition. The main contributions of this paper are as follows:(1). This paper gives a comprehensive analysis of the research background, significance and current status about sparse and low-rank representation theory and face-recogntion problems, introduces theories relate to sparse and low-rank representation, techniques of discriminant analysis, and some discrimination models based on sparse and low-rank representation in detail.(2). Inspired by the method of Fisher discriminant analysis, this paper proposed a novel algorithm, i.e. Fisher Discrimination Based Low Rank Matrix Recovery (FDLR). The introduction of discriminant regularization in algorithm FDLR promotes the discrimination power, which improves the performance of face recognition efficiently.(3). This paper applies algorithm FDLR to face recognition with the consideration of the occlusion problem existed in testing or training sample. Differing from the normal algorithms that discard the sparse error, paper further improves the recognition rates by keeping the sparse error calculated from algorithm FDLR and combining the sparse error with appropriate classification strategy.Simulation experiments on the basis of face databases (Extended Yale B、CMU Multi-PIE and AR) have been done in this paper, and the experiment results show the improved algorithm proposed in the paper can acquire preferable recognition rates, even in the case of the random picture elements of test sample has damaged in varying degrees. The algorithm proposed in the paper also has good robustness.