Research On Applications Of Sample Selection And Feature Selection In Linear Representation-based Face Recognition

Because of its wide applications in the identity authentication,public safety,business,bank and other areas,face recognition has become one of the most popular topics in the fields of computer vision and pattern recognition research since decades ago.Linear representation?LR?based methods have caught much attention of researchers among other face recognition techniques.The core idea of these methods,called linear subspace assumption,is that samples from the same class distribute in the same linear subspace.Though plenty of progress and breakthroughs on LR based face recognition have been made,it is still a worthwhile research topic.In this thesis,some improvement and exploration,in terms of sample selection and feature selection,aiming at problems of the existing LR based methods is conducted.The main works and contributions of this thesis are as follows:1.A sparse reinforced collaborative representation classification method is proposed in this thesis.Sparsity of code vectors is enforced by employing multiple ₂ norm based regularization terms which constrain code vector from different aspects.At the same time the proposed method can be regarded as a unified framework of kin methods and covers several similar collaborative representation classification methods with ₂ norm based regularizations.Also,relationship between different regularizations and corresponding correlations of training data is analyzed in this thesis.2.Classification discriminant of directly using class-wise representation residuals is weakened by the inconsistency between objects of presentation and classification.To deal with this problem,the representation discriminant criteria?RDC?is proposed in this thesis.It measures the capability of a linear representation discriminating the best class from others with ratio of the smallest class-wise representation residuals and the second smallest one.A more discriminative representation can be obtained by minimizing the RDC,and consequently a more confident classification decision could be made.3.The distribution of face images in face space may violate the linear subspace assumption due to the impacts of occlusion and illumination.Fortunately,the linear subspace assumption still holds in the subspace spanned by unaffected image features.By borrowing hypothesize-verify loop from random sample consensus and combining with RDC,face recognition under condition of occlusion and illumination is casted as a feature selection problem.First,images are projected into a subspace spanned by some randomly selected image features.And then,RDC is applied to measure discriminant of the linear representation built on the selected features.Finally,the classification is made with the most discriminant linear representation.The propose method shows high robustness against occlusion and illumination.