Face Recognition Based On Virtual Face And Sparse Representation

Face recognition is a biometric technology based on face feature information which is used for personal identification.Sparse representation-based method is an important method which has been attracted by scholars in face recognition.In face recognition,the number of samples of a data set is limited.Limited number of samples is not able to reflect the possible changes of face images.Moreover,conventional sparse representation-based classification method(CSRCM)only considers a kind of norm constraint and the time consume is high.In order to strengthen the representation capability of data set and reduce the time complexity of CSRCM,this paper will show three virtual face construction methods based on face symmetry,three improved sparse representation-based classification methods(SRCMs)and two score fusion approaches.Firstly,three virtual face construction methods design corresponding frameworks based on face symmetry to generate available virtual sample.Time complexities of three methods are different.The virtual samples generated by one method are available when the rotation of original sample is not great.They can reflect the possible changes of face image which are not included by original data set.However,when the rotation of original sample is great,the other two methods perform better.Secondly,three improved SRCMs exploit grouped sparse representation strategy to reduce the time consume and simultaneously use l1 and l2 norm constraint to obtain soluble vector.In the step of score fusion,these three methods not only consider the score from different groups,but also exploit the inner product,external product and deviation of reconstructed samples to reflect the relation between group and group.Through comparative analysis,these three methods have their own merits.The test sample is more likely to be classified to correct class by fusing the scores from different groups and corresponding relation coefficients of the inner product,external product and deviation of reconstructed samples.Finally,two score fusion approaches are adaptive.The first approach takes the test sample into account and generate optimal coefficient for each score.The second approach designs a real function and generates appropriate coefficient for each score by using the area of the real function and rectangular coordinate system.Through comparative analysis,the fusion results of the two approaches are able to better classify the test sample.Meanwhile,in the 6 chapter,this paper made an effective combination of above 8 proposed methods and compared the combinatorial algorithms with the other 6 state-of-the art methods in five face databases.Experimental results these five face databases demonstrate that the 8 proposed methods is rational and can obtain higher recognition accuracy than 6 state-of-the art methods.