정규화된 심전도를 이용한 분해능 조정된 2D 스펙트로그램 기반 사용자 인식에 대한 연구

Abstract

Recently, security technologies that prove personal identity are being upgraded to biometrics systems using bio-signals that are resistant to forgery and alteration. Research-based on user recognition systems using the representative electrocardiogram(ECG) signals are actively underway among bio-signals. To measure and recognize ECG considering the real-life environment, a recognition system using ECG signals acquired in a complex state is required. In order to apply a user recognition system using an ECG in real life, the purpose of this study is to solve the problem with decline recognition performance due to motion artifacts generated when ECG signals are acquired in a dynamic state, low recognition accuracy due to multidimensional features, and high time complexity due to complex networks.

Existing normalization methods do not consider morphological features or are complicated in the operation process. Existing multidimensional feature extraction methods are designed with low recognition accuracy and complex network structures. In this study, an adaptive similarity filter-based normalization was developed to consider individual morphological features in the preprocessing stage of the recognition system. Besides, a resolution adjusted 2D spectrogram multidimensional features by optimization were utilized as a single convolution neural network(CNN) designed with a low hidden layer structure. Accordingly, the problem with decline recognition performance caused by artifacts was solved with a similarity filter considering individual morphological features. The low recognition accuracy due to the multidimensional features was solved with a resolution-adjusted 2D spectrogram. The problem of high time complexity caused by the complex network was solved with a resized input image and a single network. Therefore, in this study, ECG in static and dynamic states was acquired and built as Chosun University database(CU-DB). This was applied with adaptive similarity filter-based normalization and resolution-adjusted 2D spectrogram. The method proposed in this study consists of a step for removing noise from an ECG signal and a step for analyzing feature data in a user recognition system.

In consideration of real life, the static states for acquiring an ECG in the CU-DB includes sitting, lying down, and sitting after exercise. The dynamic states are phone touch, door opening and closing, and stepper exercise. The performance of the proposed normalization method and multidimensional feature extraction method was compared with existing methods and analyzed using the public databases MIT-BIH normal sinus rhythm(NSR), QT, European, and Arrhythmia. The average similarity rate of the proposed normalization method was analyzed to be 4.18%, 1.36%, 6.81% higher in Euclidian distance than the existing time(cross correlation), frequency (optimized band pass filter), and phase normalization method. In addition, in the Mahalanobis distance, 2.86%, 0.96%, and 2.87% cosine similarity were analyzed to be 4.08%, 1.58%, and 5.99% higher than the existing method. Accordingly, the ECG morphological features of the static and dynamic states normalized based on the adaptive similarity filter were closer than the existing normalization method, and the recognition performance was improved by an average of 2% compared to before normalization. The problem of low accuracy was solved as the recognition performance using the proposed resolution-adjusted 2D spectrogram was analyzed 2% and 0.4% higher in CU-DB than the existing multidimensional feature extraction methods of ensemble empirical mode decomposition and Mel frequency cepstrum coefficients. Recognition performance using the proposed 2D spectrogram was analyzed as 97.1% in CU-DB, 100% in MIT-BIH NSR DB, 98.1% in QT DB, and 98.7% in European DB. This was analyzed to be 1.4% higher in the maximum European DB and 0.1% in the minimum Arrhythmia DB than the existing single-dimensional, multidimensional, and network methods. Recognition performance using a 2D spectrogram of 1/4 image size was maintained in a single network CNN and the elapsed learning time was shortened by 6 seconds.

In this thesis, a user recognition system-based study was conducted using an ECG acquired in a complex state considering real life. Because it was verified through the ECG DB in a complex state, it is expected that it can be applied using an ECG signal acquired in real life. In the future, I plan to study a biometrics system that will be applied to the driving environment for intelligent vehicles. I also plan to increase the number of training data, optimize networks, and conduct research on lightweight network design to further improve the recognition performance using ECG in deep learning-based networks.