Electrocardiogram Signal Biometric Recognition Based on Enhanced Self-Attention Mechanism

As a secure biometric recognition technology, electrocardiogram (ECG) identity recognition possesses extremely high security since it cannot be obtained or replicated in a deceased state. However, due to variations in the sampling frequencies of ECG signals, there exists inconsistency in their periods and time intervals. The current mainstream model architecture is greatly affected by this. Therefore, it is recommended to integrate ECG signal features to create a model with higher recognition accuracy. Firstly, by employing an adaptive sliding window strategy to address the issue of inconsistent periods in ECG signals, effective segmentation has been achieved to enhance the contextual information and features of electrocardiogram signals. Secondly, introducing the BlurPool structure for downsampling after the feedforward neural network helps maintain local features, reduce the size of feature maps, and preserve detailed information in the electrocardiogram. Lastly, adopting the LeakyReLU activation function and adding a linear layer handles the non-linear relationships within the electrocardiogram signals. Model has achieved a <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{0. 3 \%}$</tex> to <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{1. 4 \%}$</tex> improvement in recognition accuracy compared to existing models. In addition, the study compared the optimal segmentation length for such signals.