search

XY-boy / DRSR

[INFFUS 2023] From degrade to upgrade: Learning a self-supervised degradation guided adaptive network for blind remote sensing image super-resolution
55 stars 1 forks source link

Can patch_i and patch_j come from two different images? #1

Open Ken1256 opened 1 year ago

Ken1256 commented 1 year ago

Or the cropping has to be done in the same image? Can DRSR be used to multiple downgrade real word SR?

XY-boy commented 1 year ago

Or the cropping has to be done in the same image? Can DRSR be used to multiple downgrade real word SR?

Hi, sorry for the late reply. For Q1: Patch_i and patch_j should come from the same image. For Q2: You can finetune the encoder to generate a more practical representation of real-world degradations as it is self-supervised. Then you can apply DRSR to real-world scenes. Note that we still need supervised LR-HR pairs for model training. So it would be better if you can finetune our pre-trained model or train it from scratch on your dataset to get better performance.

Ken1256 commented 1 year ago

It is because it must has same downgrade or limit randomness? But real word image difference area have difference downgrade.

XY-boy commented 1 year ago

It is because it must has same downgrade or limit randomness? But real word image difference area have difference downgrade.

Real-world SR is a complicated issue, we simplify the degradation with this assumption. Maybe you can refer to this paper to address the spatially variant issue. Mutual Affine Network for Spatially Variant Kernel Estimation in Blind Image Super-Resolution