DUAL-DOMAIN LOW-RANK FUSION DEEP METRIC LEARNING FOR OFF-THE-PERSON ECG BIOMETRICS

Electrocardiogram (ECG) biometrics has been an emerging field, and off-the-person ECG biometrics capturing the ECG from fingertips is one of the new trends in this field. However, dynamic morphological variability in the same person and low signal-to-noise ratios pose great challenges for off-the-person ECG biometrics. To reduce the dynamic morphological variability, this paper introduces deep metric learning into ECG biometrics to learn intra-individual compact features. To enforce the robust of proposed method, dual-domain features extracted from both 1D signals and 2D spectrograms are integrated by low-rank fusion. Furthermore, this method dispenses with the need for noise removal and outliers discarding completely. Experiments on two off-the-person ECG benchmark databases demonstrate that the proposed method significantly outperforms the state-of-the-art methods. Additionally, ablation experiments show the effectiveness of every part of our framework.