Denoising final
Download as PPTX, PDF0 likes84 views
This document discusses using a generative adversarial network (GAN) to remove muscle noise from electrocardiogram (ECG) signals. It provides background on ECGs and current denoising methods. PhysioNet is introduced as a source of ECG data, while the CapgMyo database provides muscle noise samples. GANs are explained as a way to filter noise. Results show the GAN improves signal-to-noise ratio over the noisy signal and achieves higher recognition accuracy than an autoencoder.
Denoising final
- 1. Generative Adversarial Network for muscle noise removal in ECG Olena Shevchenko Eleks
- 2. Introduction
- 8. Performance Study of Different Denoising Methods for ECG Signals Source: https://www.sciencedirect.com/science/article/pii/S1877050914010138 Mohammed AlMahamdya , H. Bryan Riley PhD
- 10. Why it works in real-world? Universal approximation theorem, Cybenko (1989)
- 11. CG signal enhancement based on improved denoising auto-encod Source: https://www.sciencedirect.com/science/article/abs/pii/S0952197616300331 Xiong et al., 2014
- 13. Data
- 14. What about Data? • Small amount • Highly personal • Research purposes only Problem: • Small amount • Highly personal • Research purposes only
- 15. What about Data? Problem: Solution: • PhysioNet • Make your own dataset • Small amount • Highly personal • Research purposes only • Small amount • Highly personal • Research purposes only
- 16. PhysioNet
- 17. Source: https://www.physionet.org/physiobank/database/nsrdb/ The MIT-BIH Normal Sinus Rhythm Database
- 18. The ECG-ID Database Source: https://www.physionet.org/physiobank/database/ecgiddb/
- 19. We need some noise!
- 20. CapgMyo: A High Density Surface Electromyography Database for Gesture Recognition Source: http://zju-capg.org/myo/data/
- 21. CapgMyo: A High Density Surface Electromyography Database for Gesture Recognition Source: http://zju-capg.org/myo/data/
- 22. CapgMyo: A High Density Surface Electromyography Database for Gesture Recognition Source: http://zju-capg.org/myo/data/
- 23. Data Samples
- 24. Let’s talk about GANs!
- 25. Generative Adversarial Nets Ian Goodfellow et al., 2014 Source: https://deeplearning4j.org/generative-adversarial-network
- 27. Results
- 28. Signal-to-Noise ratio As a metrics of denoising, we use Signal-to-noise ratio (SNR). SNR is a measure that compares a level of the desired signal with a level of background noise. Where X - is original, Xm - is noisy. User 2 34 35 52 72 Average SNR noisy signal (dB) 5,06 0,77 3,81 0,80 4,50 3,99 SNR filtered with GAN (dB) 20,26 20,33 20,01 18,71 20,70 21,41
- 29. User Clean signal from chest Noisy signal from hands GAN filtered signal AE filtered signal 2 96,67% 69,46% 92,04% 94,30% 34 98,28% 91,29% 99,03% 98,60% 35 99,57% 90,75% 98,71% 96,67% 52 99,14% 71,18% 97,31% 96,67% 72 97,31% 70,54% 95,38% 94,30% Average 98,53% 86,38% 97,24% 96,82% Identity Recognition accuracy of SVM classifier and 1 vs 100 users setup. Signal to Noise (SNR) ratio results for noisy and filtered signals. Few identities from ECG are highlighted along with the overall average values. Results
- 30. Thank you! LinkedIn: Helena Shevchenko olenashevch@gmail.com