Illumination Robust Face Recognition Using Low-rank Matrix Decomposition

Face recognition technology as a kind of non-contact automatic identification technology, is widely used in the field of public security, entrance guard control, entry and exit management and so on. It has great potential application value. At present, Face recognition has been successfully applied in real life under the condition of certain restrictions. But when the restricted conditions are not met, the recognition performance will decline substantially. Therefore, unconstrained facial recognition technology research has become the most concern of the current face recognition field. Unconstrained conditions contain illumination, shade, fuzzy, posture and other external conditions, which are important factors resulting in a decline in face recognition performance. Illumination as a key factor is a hot spot in research of face recognition. How to effectively get the illumination robust features is the key to improve the recognition performance under illumination change condition.Based on the analysis of the status quo of domestic and international research methods to solve the illumination problem of face recognition, we locate our study on the illumination robust feature extraction based on low-rank decomposition theory, according to the development tendency of correlation theory. In this paper, we study the theory of low-rank decomposition and verified its performance of illumination problem. The result of low-rank decomposition is poor under uneven illumination distribution and contains illumination component. For this problem, we propose a low-rank relative gradient histogram feature extraction method combining the characteristic that relative gradient image is robust to illumination. The method applies low-rank decomposition on relative gradient magnitude images instead of the original images, we get the image which is more robust to drastic illumination. We extract low-rank relative gradient histogram feature after consulting the implementation of Patterns of Oriented Edge Magnitudes. The feature preserves the local edge texture and neighborhood information, it effectively improves the performance of face recognition under illumination change condition The main work can be summarized as follows in our study paper:①based on the research of human face illumination processing method in domestic and international, we further study the low-rank decomposition theory and its typical algorithm, then analyze the low-rank decomposition principle and the implementation process of each algorithm in detail, we compare decomposition performance of these algorithms.②face recognition method based on the low-rank decomposition is studied, first we study the sparse representation and its implementation algorithm which used to identify classification. Then we describe the process of face recognition method based on the joint of low-rank decomposition and sparse representation and compare this method with the method based on sparse representation on illumination face database.③we discuss the gradient and the relative gradient of face images, we get the low-rank relative gradient image by doing low-rank decomposition on relative gradient images. Combining feature extraction method of Patterns of Oriented Edge Magnitudes, we extract the feature more robust to illumination. The experiments are carried out on FERET, Yale B and PIE, the comparisons of the experimental results verify the effectiveness of the proposed method.Combining advantages of low-rank decomposition, relative gradient and Patterns of Oriented Edge Magnitudes, we propose a Low-rank Relative Gradient Histogram Feature description method. Compared with Relative Gradient Histogram Feature, Patterns of Oriented Edge Magnitudes and low-rank feature method, this proposed method is robust to illumination changes, especially drastic illumination.