Use - Githubissues

cszn / DPSR

Deep Plug-and-Play Super-Resolution for Arbitrary Blur Kernels (CVPR, 2019) (PyTorch)

https://cszn.github.io/

837 stars 211 forks source link

Use #2

Open victorca25 opened 5 years ago

victorca25 commented 5 years ago

How can I test with my own images? It's not clear to me how it's useful without the matlab files

wuxiaolianggit commented 5 years ago

@victorca25 do you have solved the problem?

victorca25 commented 5 years ago

@wuxiaolianggit nope

MollyDai commented 5 years ago

I also have this question, what does non-blind SISR mean?

lucasjinreal commented 5 years ago

@MollyDai @victorca25 I think this methods need label to predict, which means it's just overfit... without training labels (kernel file), the blur image not exist in training set result comes very bad

alsombra commented 5 years ago

@MollyDai Non-blind SISR algorithms assume you already know the blur kernel, and they normally use this information for doing the Super-Resolution step. To use a non-blind SISR model you need to first use some kind of kernel estimation method to determine the blur kernel.

yuzefang96 commented 5 years ago

Is there any kernel estimation code for us to generate our own .mat file?

victorca25 commented 5 years ago

Or any alternative to the .mat files

magneter commented 5 years ago

How could I generate the .mat files of my own images?

yuzefang96 commented 5 years ago

I've found some kernel estimation codes but they need both the original image and the blurred image. What if i only have the low quality image?

oobbppoo commented 5 years ago

Is there any kernel estimation code for us to generate our own .mat file?

I've found some kernel estimation codes: https://github.com/rgbitx/image_deblur_code

alsombra commented 5 years ago

I have not yet fully understood the project. I have to look more throughly and carefully into it. But I found that the degradation kernels are given. There are commented instructions on the code saying that you can generate your own kernels OR you can download the .mat files (kernels) here.

lucasjinreal commented 5 years ago

@alsombra In short, the author just using training samples (exactly training ones) to predict result.

Which we can also called: overfit.

You can not predict on a random natural blurred image and rescale it up.

cszn commented 5 years ago

DPSR is non-blind model-based SISR method which assumes the blur kernel is known.
Blur kernel estimation methods: https://github.com/rgbitx/image_deblur_code
Motion blur kernel generation code: http://home.deib.polimi.it/boracchi/Projects/PSFGeneration.html
We did not even use any blurry LR images for training.
For the kernel sensitivity of SISR methods, please refer to Accurate Blur Models vs. Image Priors in Single Image Super-Resolution.

yuanzhenjie commented 5 years ago

I have not yet fully understood the project. I have to look more throughly and carefully into it. But I found that the degradation kernels are given. There are commented instructions on the code saying that you can generate your own kernels OR you can download the .mat files (kernels) here.

I download it, these kernels are only needed for testing. I think.