search

shadyabh / Correction-Filter

76 stars 8 forks source link

PSNR results #6

Closed FuzhiYang closed 3 years ago

FuzhiYang commented 4 years ago

When I evaluated PSNR on Set14 dataset. I blured the image with 32x32 gaussian blur kernel with sigma 2.5/sqrt(2), and then I used matlab's bicubic imresize function to x2 downsample it like below: comic

Then I ran "correct_imgs.py" to get the corrected results like this: comic_x2_corr_corrected

Finally I forward this corrected image to RCAN_x2 model and got this: comic_x2_corr_corrected

But the PSNR on Set14 is only about 24, which is far lower than the reported number in the paper, 31.272. Did I make some mistakes during these steps? Could you share the details of your degradation process? (gaussian blur, bicubic downsample, etc)

shadyabh commented 4 years ago

This is the wrong way to downsample an image, you need to blur then subsample not blur then bicubic downsample. First, you apply 32x32 Gaussian then you will need to subsample it directly or work with a stride in your convolution. Please check "downsample_using_h" in utils.py.

Good luck