Learning the Degradation Distribution for Blind Image Super-Resolution�CVPR 2022

Jeongho Min

M.S Student, UNIST

7^th October 2022

Blind SR

EO Image

SAR Image

https://slideplayer.com/slide/17037100/

Image Super Resolution

https://medium.com/@hirotoschwert/introduction-to-deep-super-resolution-c052d84ce8cf

Paired Images

Previous Works in SR

1. Predefined Degradation-based��e.g.) Bicubic downsampling, multiple blur kernels and random noises, etc.�Use larger degradation space ! => fail in real applications�
2. Degradation Learning-based��learn the degradation adaptively with adversarial training�e.g) CycleGAN frameworks => deterministic model , fail to model random factors in degradation��Use “Degradation Pool” contains blur kernels and noises from the real-world images!

Using GAN

To learn image super-resolution, use a GAN to learn how to do image degradation first(ECCV 2018)

Overview

PDM(Probabilistic Degradation Model)

Learning the Degradation Distribution

Kernel Module

Noise Module

Degradation distribution :

Degradation process :

k: size of the blur kernel

Learning the Degradation Distribution

Degradation distribution :

Degradation process :

PDM (Probabilistic Degradation Model)

PDM Total Loss Function

Overview

PDM(Probabilistic Degradation Model)

Learning the Degradation Distribution

Degradation distribution :

Degradation process :

PDM (Probabilistic Degradation Model)

PDM Total Loss Function

Contributions

1. PDM is able to model more diverse degradations�: allows one HR image => degrade into multiple LR images, more training samples bridge the gap between training and test datasets.
2. Priori knowledge about the degradations => learn degradation better�: reduce learning space of PDM and easier to be trained
3. PDM degradation process => linear function �: better decouples the degradation with image content
4. PDM is better constrained and easier to be trained �: possible to train PDM and the SR model simultaneously

​

​

Experiments

Experiments

Experiments

Q&A