Sparse Representation For Face Recognition Based On Fisher Discrimination Dictionary

The theory of sparse representation breaks through traditional limitation of Nyquist sampling theorem.It is one of the most brilliant achievements in the area of signal processing.Because of its excellent classification performance and the robustness of noise and occlusion,sparse representation is paid close attention to.It is successfully applied to face recognition in 2009,providing a new method for the research of face recognition.Dictionary learning has long been a research area of sparse representation to face recognition.It can increase the accuracy in face recognition to learn a dictionary of high-performance.Therefore,varieties of dictionary methods came up.Among all the methods,Fisher Discrimination Dictionary Learning brought Fisher discrimination criterion dictionary learning pattern.Then developed into structured dictionary with dictionary recognition ability and coding coefficient recognition ability.After that,test samples were sparse representation classified on the dictionary.Thus increase the accuracy in face recognition remarkably.To make further efforts of increasing the accuracy of algorithm in recognition,this thesis takes dictionary learning and discrimination of sparse coding coefficient into account,studies methods based on Fisher Discrimination Dictionary Learning,and proposes two improving methods.Research work and innovations in this thesis are as follows:(1)A locality-constrained group sparse representation method for face recognition based on FDDL was proposed in this thesis.In order to get more discriminant coefficients,training sample basing on FDDL algorithm should be learned first,then more discriminant dictionaries and coefficients are got.After that,the test sample is sparse represented by the learned dictionary.In the process of sparse coding,the dual constraints of locality and group sparsity are applied,and the local structure information between the test sample and the nearest neighbors of dictionary atoms is preserved,at the same time,the test sample can be represented with few dictionary categories.Finally,test samples are classified by discriminant representation coefficient in phases of training and testing.Contrast experiments on common face databases verify that this method can increase the accuracy of face recognition.(2)A joint sparse representation method for face recognition based on FDDL was proposed in this thesis.In order to make full use of label information of samples,to obtain the sparsity and discrimination of coding coefficient,the joint sparse constraint is applied in the sparse coding process of the test sample on the learned dictionary.That is to say,the test sample is sparse represented in class-specified sub-dictionary,the category of the test sample is predicted by representation residuals.Then corresponding groups are weighted constrained by forecasting value.When choosing class-specified sub-dictionaries to represent,we choose them as few as possible and those closer to test samples,which can increase discrimination of coding coefficient;Combining global sparse method,the sparsity within selected group is guaranteed.Finally,test samples are classified by discriminant representation coefficient in phases of training and testing.Experiments results on common face databases verify the validity of this method.