The Research On Sparse Representation-Based Face Recognition Methods

Face recognition is an important biometric identification technology which has wide application prospect. In this research, multiple research fields such as mathematics, picture processing, pattern recognition, computer vision are involved. In recent years, sparse representation has been successfully applied to face recognition. The sparse representation-based face recognition methods are highly robust to the light condition, the variety of facial expression, and partial occlusions, and this advantage helps it attract many research interests. Currently, the research mainly focuses on the structure of the redundant dictionary, sparse reconstruction algorithms, classification methods, and so on. The sparse representation-based face recognition methods are studied in this paper. To improve the recognition rate, we designed new dictionary learning algorithms. Firstly, a new constraint is incorporated to the sparse coding objective function in order to improve the discrimination of dictionary. Secondly, the dictionary learning algorithm is extended for color face recognition, so that the color representation ability of the learned dictionary can be improved. The main contributions of this paper are listed as follows:1?To overcome the drawbacks of metaface learning (MFL) algorithm, an algorithm called coefficient-similarity-based metaface learning (CS-MFL) is proposed. In CS-MFL, the coefficient similarity is incorporated as a new constraint to the original sparse coding objective function when updating the sparse representation coefficient, utilizing the similarity among the training samples from the same person to make the learned dictionary more discriminative. To solve the new optimization problem, both l2 norm-based constraints are combined, and the original problem becomes a typical l2-l1 problem. Experiment was carried out on Extended Yale B, AR and ORL face databases, the results show that the proposed CS-MFL algorithm can achieve higher recognition rate than MFL algorithm, comparing with MFL, the maximum increments of recognition rate are 0.58%,1.15% and 2.5%, which demonstrates that the dictionary learned by CS-MFL algorithm is more efficient and more discriminative than the traditional MFL for face recognition application.2?In order to utilize the correlation among different channels, basing on the label consistent K-SVD (LC-KSVD) algorithm, a new dictionary learning method for color face recognition (CE-LC-KSVD) is proposed. To improve the representing ability of each atom for color images, this algorithm concatenates R, G and B values into a single vector, and then introduces a new inner product into orthogonal matching pursuit (OMP) during sparse coding procedure. The experiment results on different color face images databases show that compared with LC-KSVD algorithm, the proposed CE1-LC-KSVD algorithm which only uses the R, G and B values can get about 7% increment of recognition rate; CE2-LC-KSVD algorithm which simultaneously adopts the new inner product of OMP can further improve the recognition rate by 1.7%-2.8% on AR databases where strong saturation exists in each face image.