Processing & Properties of Floor and Wall Tiles.pptx
Noise2Score: Tweedie's Approach to Self-Supervised Image Denoising
1. Noise2Score: Tweedie’s Approach to Self-Supervised Imag
e Denoising without Clean Images
Kwanyoung Kim, Jong Chul Ye
Bio Imaging, Signal Processing & Learning Lab
@KAIST
2. Supervised Learning Approaches
• CNN as a direct inverse operation
• Most simplest and fastest method
• Supervised learning with lots of data
: feed-forward network
6. Self-Supervised Learning: SURE
Soltanayev et al, NeurIPS, 2018
Self-supervised denoising using Stein Unsupervised Risk Estimator (SURE)
Divergence-based
penalty
Autoencoder loss
7. Supervised learning, Noise2X, SURE
Supervised learning
Noise2X : Noise2Noise, Noise2Self, Noise2Void, etc.
:samples of the noisy-clean image pair for training data
: target z is related to in unique ways depending on algorithms.
most of the algorithmic choices are heuristic
SURE
: divergence penalty is added to compensate for use of y
10. Tweedie’s formula for general exponential family
• Using the Bayes’ rule, the posterior density:
• Probability distribution of exponential family:
B Efron, Journal of the American Statistical Association, 2011
11. Tweedie’s formula for general exponential family
• The closed form solution for the posterior mean:
• Bayes optimal solution posterior mean
12. Case1. Gaussian noise
• Tweedie’s formula calculates the posterior mean of x given y.
• Gaussian noise removal
Score function
16. Tweedie’s formula of exponential family distribution for
image denoising
As long as we can compute the score function,
optimal denoising can be achieved by using Tweedie’s formula.
22. Denoising Autoencoder (DAE)
• DAE is to learn to recover data from the perturbed data
• DAE can be used to estimate the score function of data
Alain et al, JMLR, 2014
Equal to Tweedie’s formula for Gaussian noises
23. Amortized – Residual DAE (AR-DAE)
• The residual from of the DAE:
• Addresses instability and reduce approximation error
• Directly estimate the score function:
Lim et al, ICML, 2020
24. Relation to SURE
By using AR-DAE transform
Score matching cost by Hyvärinen et al, JMLR, 2005
30. Conclusion
• Noise2Score: a novel unified framework for self-supervised image
denoising.
• Our Noise2Score can be applied to any image denoising problem from
exponential family noises.
• Identical neural network training can be used regardless of noise model.
31. Thank you!
Jong Chul Ye
E-mail:
jong.ye@kaist.ac.kr
Kwanyoung Kim
E-mail:
cubeyoung@kaist.ac.kr
Editor's Notes
그런데 지금까지의 연구는 대부분의 경우 과화질의 레이블의 영상이 존재화고 저화질의 측정 데이타가 있을떄 이를 지도학습의 기반으로 복원하는것이었는데, 만약에 고화질의 영상이 존재하지 않을경우는 어떤게 딥네트워크를 훈련하여 사용할수 있을까요?
그런데 지금까지의 연구는 대부분의 경우 과화질의 레이블의 영상이 존재화고 저화질의 측정 데이타가 있을떄 이를 지도학습의 기반으로 복원하는것이었는데, 만약에 고화질의 영상이 존재하지 않을경우는 어떤게 딥네트워크를 훈련하여 사용할수 있을까요?
지금까지 본 발표에서는 GAN이 의료영상 복원에서 비지도 학습기법으로 점점도 중요한 주제가 되고 있다는것을 보였고, 특히 collaGAN은 MR contrast 의 연구에 사용이 가능하다는것을 보였습니다. 경청해주셔서 감사합닏.