Search for a command to run...
Electrocardiogram (ECG) has become a popular biometric to study since it is highly secured against spoofing attack. In this study, we address the issues of hard-required ECG data length and neglected causality when performing ECG identity matching tasks. First, we propose an ECG image generation algorithm that is able to handle any specified number of ECG heartbeats. Such an algorithm uses detected R-peaks as folding points and projects ECG data onto a two-dimensional image, which overcomes the challenge of hardly-required fixed length and truncated ECG. Second, we leverage transfer learning and perform across- session testing. We construct the ECG identification models based on the pretrained AlexNet and ReseNet18 models. Our ECG biometric models are trained on the past ECG data and their performances are evaluated on future ECG data. Furthermore, we develop a voting strategy that is able to detect anomaly ECG heartbeats. Our novel ECG image generation approach shows itself to be competitive. Such method has been evaluated on the MIT-DB and ECG-ID datasets. We observe satisfying results of the proposed models in both datasets: 100% on the MIT-DB and 94.4% on ECG-ID. More importantly, our method is available to generate satisfying results by using a single ECG beat to conduct identity matching task: 100% on the MIT- DB and 91.7% on ECG-ID. In addition, qualitative analysis presents the perceptual uniqueness ofECG between individuals. We believe that the proposed ECG biometric system is promising to identify humans with short ECG sequence.