Non-fiducial based electrocardiogram biometrics with kernel methods
Electrocardiogram (ECG) biometrics is a relatively novel trend in the field of biometric recognition. ECG is a new generation of biometric modality which presents a number of notable problems in signal processing, extraction of significant features from the ECG signals and construction of an accurat...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English |
| Published: |
2017
|
| Online Access: | http://psasir.upm.edu.my/id/eprint/69992/1/FK%202017%2091%20-%20IR.pdf http://psasir.upm.edu.my/id/eprint/69992/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Electrocardiogram (ECG) biometrics is a relatively novel trend in the field of biometric recognition. ECG is a new generation of biometric modality which presents a number of notable problems in signal processing, extraction of significant features from the ECG signals and construction of an accurate subject recognition system. These notable problems are due to time-varying nature of ECG signals implying cardiac conditions and the type of ECG signal acquisition. This thesis considers all the inherent processes to an ECG biometric system involving pre-processing, feature extraction and classification. The thesis proposes a novel ECG verification technique based on non-fiducial approach which explores waveform itself using kernel methods for feature extraction and classification after preprocessing (denoising ECG signals) one lead ECG signals of 52 subjects. For ECG signal processing, Coiflet3 wavelet and Rigrsure rule of hard threshold is proposed after evaluating different discrete wavelets based on statistical measuring criteria which include cross-correlation, signal-to-noise ratio, reconstruction error, root-mean-square error, and others. A new non-fiducial approach is proposed for feature extraction. This approach constructs an algorithm by combining autocorrelation (AC) and Kernel Principal Component Analysis (KPCA) techniques. The effectiveness of this algorithm is investigated by comparing with other AC based feature extraction algorithms involving AC/LDA (Linear Discriminant Analysis) and AC/PCA (Principal Component Analysis). At classification level, Gaussian multi-class Support Vector Machine (SVM) with the One-Against-All (OAA) approach is proposed to evaluate verification performance rates of the feature extraction algorithms. The results of analysis demonstrate that the AC/KPCA has a maximum effect on achieving high subject and window recognition rates in different operational conditions. The highest window and subject predictive accuracies achieved are approximately 92% and 77% on KPCA data set with the lowest biometric error and overfitting. The lowest biometric errors of false non-match rate and false match rate are decreased to about 6.19% and 1.79%, respectively on the KPCA data set. |
|---|