Biometrická identifikace na základě klasifikace EKG signálů

Abstract

This thesis deals with the possibility of using ECG signals as identifiers for biometric identification and authentication systems. In this thesis, methods and processes of biometric identification are described in detail with a strong focus on the necessary requirements that such systems and biological identifiers must meet. Also included is an extensive survey of recent scientific results, methods, and procedures in this area. As part of this work, classification systems based on pre-trained 2D convolutional neural networks have been developed and tested. These systems were tested on a large database of binary images and STFT spectrograms generated from publicly available databases of ECG signals designed for biometric identification purposes. In this work, it was found that binary images exhibit significantly higher classification success rates than the nowadays very popular time-frequency analysis images, specifically those created using the Fast Fourier Transform. Furthermore, the effects of various parameters and features of ECG signals on the overall classification success rate, especially the effect of temporal variability of ECG signals, were described in detail and partially tested in this work. In this testing, a clear negative influence of the temporal variability of ECG signals on the overall classification success rate was confirmed, especially for STFT spectrograms. The overall results of this work confirm the sufficient uniqueness of ECG signals and the applicability of biometric identification systems based on them for routine practice.