search

sefibk / KernelGAN

Other
337 stars 77 forks source link

Question about kernel estimation #56

Closed rptxk01 closed 3 years ago

rptxk01 commented 3 years ago

Hi, thank you for the great work. I have a question regarding the paper.

KernelGAN estimates down-scale kernel from LR and considers that kernel to be the kernel of unknown clean HR.

I am wondering if there are domain-related problems in the process.

For example, if LR is highly degraded by blur or compression, the estimated kernel will map "degraded LR" → "degraded down-scaled LR". However, the ideal kernel should be "clean HR" → "degraded LR" in this case.

I think the domain mismatch may deteriorate the restored SR quality.

Could you enlighten me on this concern? Thanks a lot.

sefibk commented 3 years ago

The "non-formal" answer is that the SR task is different than the deblurring and denoising task - it generally deals with clean images that were downscaled and "lightly" downgraded. It doesn't deal with severe degradations. Regarding deblurring and denoising, we had an idea to apply KernelGAN for that however it is not straight forward and didn't work for us with our first try.

rptxk01 commented 3 years ago

Thanks for the detailed replies. The comments inspired me a lot.