Research On ECG Biometrics Based On Sparse Representation And Dictionary Learning

In recent years,after face recognition,fingerprint recognition,voice recognition,movement recognition and other biometrics,electrocardiogram(ECG)recognition,with its unique advantages(i.e.,vivo detection,high privacy and high security),has become a new recognition technology,which obtains much attention.Although ECG signal has been successfully applied in biometrics,its performance is not as good as that of other biometric traits.The main reasons are that ECG signal is often affected by various noises,causing poor stability,and its discriminative features have not been fully extracted.In order to improve the robustness and recognition performance of ECG recognition,it is necessary to study new methods to overcome the above problems.In this thesis,non-fiducial features of ECG signal are extracted as original features,and then new features are extracted from two perspectives.In order to solve the sensitiveness of ECG recognition methods to noises and large intraclass changes,the first work is developed from the perspective of multi-feature fusion,in which a multi-feature fusion method is proposed for ECG recognition.Firstly,the method takes advantage of the idea that training samples have local invariance in the inherent structure of data distribution and locality leads to sparsity,local constraint is added into sparse coding to make that the multi-feature data of the same individual can make similarity contribution to recognition.Then,the method adds a low-rank constraint for the dictionary to ensure that noised ECG can recover effectively.Finally,the cross-feature regularizer is used to enhance the smoothness of multi-feature data to explore the geometric structure of cross-feature and mine advantages of various features.Experimental results on MITDB and ECG-ID databases show that,the proposed method achieves more discriminant features and better recognition performance,which improves the robustness of ECG recognition.In order to solve that ECG recognition based on the traditional sparse representation and dictionary learning does not have power to deal with large intra-class changes and small interclass changes,and these methods have high time cost,this thesis proposes a label-guided dictionary pair learning method for ECG biometric recognition from the perspective of feature learning.Firstly,the sparse coefficients with lp norm constraint are avoided by replacing the original single dictionary by dictionary pair.The dictionary pair include a projective dictionary and a reconstruction dictionary.The projective dictionary is used to project original data into new subspace and achieve signal representation.The reconstruction dictionary is used to reconstruct the projected data.Then,the Fisher-like regularization term is added for the projective dictionary to increase inter-class scattering and decrease intra-class scattering.At the same time,in the dictionary learning,the introduction of label information of each dictionary item improves the discriminability.Experimental results on MITDB and PTBDB databases show that,the dictionary pair not only improves the recognition efficiency,but also brings superior recognition accuracy.