Convolutional Neural Networks-Based Model Architecture for Signal Processing Applications
Loading...
Date
2020
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
FACULTY : Mathematics And Computer Science - DEPARTMENT : Computer Science
Abstract
With the rapid development of biometric identification in healthcare, the bio signal of human being (ECG, EEG, EMG) has attracted much attention from the research community. However, the raw signal acquired by different sensors mostly is noised and consequently affects the biometric system accuracy. Unlike the handcraft features algorithms, deep learning algorithms have proved their ability to learn automatically the original data without mostly any preprocessing. Among of those, CNN is widely used ECG application and achieved a remarkable result. However, the exiting CNN architectures are mostly conducted for 2D data. To deal with this challenge. Deep learning methods such as algorithms based on Convolution Neural Networks (CNN) can be used to avoid conventional extraction crafting of features from ECG signals. We have designed two CNN architectures including 1D and 2D ECG arrhythmia classification. On this experiences, various fine tuning hyperparameter training strategies based on various database were implemented and tested. All results have achieved satisfactory result. In case of ECG based biometric, we have proposed a new variant of CNN called LBCNN, which is carried out on well-known databases and without any pre-processing step. An equal error rate of 0.5 % via LBCNN was achieved using ECG spectrogram energy. Lower EER with LBCNN proves that 1D LBCNN can be applied for ECG detection and classification.
Description
Keywords
ECG, biometric, CNN, LBCNN, Deep learning.