Structural sparse representation with class-specific dictionary for ECG biometric recognition

Abstract

For traditional person recognition, most methods are knowledge-based and token-based such as the methods using password, ID number or other secret information, and to enhance its security, biometric recognition is widely used in many areas. As a biometric trait, electrocardiogram (ECG) signal is gaining more attention due to the fact that it is not easy to be faked compared with other traits and only exists in living subjects. However, ECG signals are influenced by numerous factors easily and have low stability, resulting in unsatisfied recognition performance. In this paper, we propose a structural sparse representation algorithm for ECG signals and learn class-specific dictionary for each class to address these problems. Firstly, we segment heartbeats for each subject’s signal. Moreover, the mixed regularization l2,1 norm is utilized for ECG heartbeats to obtain a robust and stable feature. It is encouraged that if the atom is used to describe the heartbeat of one subject’s heartbeat, it may be used to describe other heartbeats of the same subject. In addition, a graph regularization is used to minimize the difference between heartbeats of intra-class. Finally, we conduct experiments on three public ECG databases to verify the proposed method and the subject recognition accuracy can reach to 99.66% on PTB database. From the experimental result, the proposed method can take full advantage of ECG structural characteristic and get a better performance compared with state-of-the-art methods.