Discriminative Structured Dictionary Learning For Face Recognition

The human face is the most critical biological feature used to convey identity and express emotions in daily communication.With the rapid development of computer technology,face recognition has gradually become the most important research direction in the field of computer vision and biometrics due to its non-contact,non-invasive and high accuracy,which has been widely used in security,surveillance,entertainment,etc.Relying on good generalization ability and robustness to occlusion,the study of sparse signal representation and dictionary learning have been rapidly developed since 2009 and have gradually become a research hotspot in the field of image classification,image denoising,biometrics,etc.Therefore,further study of face recognition algorithm based on dictionary learning for the purpose of improvement in discrimination,generalization ability and classification accuray,not only have a very high scientific significance,also have a very important practical significance.Based on the existing dictionary learning algorithms for classification,two discriminative dictionary learning algorithms for face recognition are proposed in this thesis.Extensive experiments on two commonly used data sets are conducted,and the results show that the proposed approach can outperform lots of recently presented dictionary algorithms on both recognition accuracy and computational efficiency.Firstly,a discriminative dictionary learning algorithm for face recognition is proposed in this thesis.It combines the category information of training samples to learn a structured dictionary,which consists of label-particular atoms corresponding to some class and shared atoms commonly used by all classes.In order to improve the discriminative ability of the dictionary and the classification accuracy,we propose to add additional label constraints to the loss function to constrain large coefficient appearance at other label-particular atoms rather than its closely associated ones,thereby expanding the differences between different categories.In addition,inspired by Laplacian Eigenmap,the Laplacian group regularization is proposed to improve the similarity of the representations for the same class and promote the label consistency.Without using?₀-norm and?₁-norm for coding regularization,we employ a simpler coding regularization?₂-norm to obtain an analytical solution and accelerate the learning process.Finally,two simple classifiers are developed to cooperate with the learnt dictionary for image recognition and they can often bring out promising results.The face recognition results show that the proposed algorithm achieves high classification accuracy and is superior to many existing face recognition algorithms.Secondly,an improved discriminative dictionary learning algorithm is proposed based on the above algorithm which combines the tradeoffs between accuracy and computational complexity.A cubic fitting model is adopted to measure the time cost of such algorithms in dictionary,and a method based on SVD is utilized to estimate the residuals in signal reconstruction.We combine the tradeoff between accuracy and computational complexity with the dictionary learning algorithm proposed above.As is shown in the experiments,the learning process become more reasonable due to the decreasing of computation redundancy.Furthermore,the results in the table compared our method to others,such as SRC,LC-KSVD,COPAR,and the basic GCC,also demonstrate the advantages of our methods in computation reduction.