In super-resolution, instead of taking loss that making high-resolution image from low-resolution image, they are looking for a latent space that would result in the same image when reverting to a lower resolution after a getting higher resolution from the pre-trained model. A method that does not require learning a large network and is not tied to network choices.
TL;DR
In super-resolution, instead of taking loss that making high-resolution image from low-resolution image, they are looking for a latent space that would result in the same image when reverting to a lower resolution after a getting higher resolution from the pre-trained model. A method that does not require learning a large network and is not tied to network choices.
Why it matters:
Paper URL
https://arxiv.org/abs/2003.03808
Submission Dates(yyyy/mm/dd)
Authors and institutions
Methods
Results
Comments