The Research Of Sparse Representation Based On Low-Rank And Eigenface For Face Recognition

As an important application of modern biometric identification technology in daily life, face recognition uses visual features of face images to identity recognition and has widely application prospects. Compared with the traditional identification methods, face information is difficult to imitate or forge and carried easily, therefore, has strong security, privacy and operational ease which make it applied in daily life. In addition, face recognition technology involving in machine learning, pattern recognition and data mining, has very high theoretical research value as an instance of interdisciplinary in life.The method based on sparse representation is a popular method for face recognition. It presents test samples as a coefficient vectors of dictionary atoms and identifies them with high accuracy and strong robustness according to the position of non-zeros number in coefficient vectors. Sparse representation theory suggests that a picture can be sparse reconstructed by other samples. Ideally, coefficients of heterogeneous samples are zeroes while similar samples are non-zeroes. In order to improve performance of sparse representation algorithm and increase the dictionary expression ability, training a dictionary has becoming a hot topic for researchers.In this thesis, we improve the sparse representation algorithm based on eigenfaces extraction and low rank representation theory, and propose an improved sparse representation algorithm based on the low rank and eigenfaces for face recognition. Practical works of the thesis are as follows:(1) We use Robust PC A algorithm (for single class images) and LRR algorithm (for multi-classes images) to extract low-rank images from original data for removing adverse effects such as illumination and expression change, so that the extracted features are more true.(2) We extract histogram of gradient(HOG) feature which has strong discrimination by capturing facial contour and texture information of face images.(3) The singular value decomposition(SVD) is used to extract eigenfaces matrix on the basis of HOG feature matrix. Eigenfaces are combined to construct a compact sparse representation dictionary for reducing the time complexity.The experimental results show that the improved algorithm proposed in this thesis has higher recognition rate compared with other classic algorithms such as SVM, SRC on several popular face datasets such as the Extended Yale B, ORL, AR and CMU_PIE. In addition, by increasing random noise and random block occlusion or real disgust on the images, the improved method shows strong robustness.