A Heterogeneous Dual-Branch Enhancement Network Architecture and a Universal ECG Spectrogram for Single-Lead ECG-Based Biometric Recognition

The electrocardiogram (ECG) signal has emerged as a promising biometric identification signal due to its inherent resistance to theft and forgery. However, due to limitations in the positioning precision of existing methods for locating waveforms and segments in ECG signals, current research struggles to utilize these waveforms and their associated prior knowledge in the recognition process. Additionally, standard convolutional or Transformer networks are usually only capable of extracting single local features or global representations, which hinders the understanding of individual differences between ECGs from various perspectives. These factors negatively affect recognition performance. To address these issues, we first introduce the PreM spectrogram, a novel universal spectrogram that can differentially represent the components of ECG signals in the time–frequency domain and achieve a similar effect without relying on precise positioning of those waveforms. Second, we introduce a novel heterogeneous dual-branch enhancement network architecture, CNN-enhanced Swin Transformer network (CESTNet), which can aggregate different features. CESTNet combines a convolutional network and a specially configured Swin Transformer to achieve identical network structures and hierarchical feature representations for extracting various feature sets from ECG signals, and through a feature fusion mechanism at each network layer, different features are fused and jointly propagated backward through the next layer of the network together. The experimental results demonstrate that our proposed method achieves 99.90%, 99.99%, 99.45% (99.57%), and 97.72% identification accuracy on the ECG-ID, PTB, CYBHi, and Heartprint datasets, respectively, outperforming other state-of-the-art methods.