[Super resolution] image Resizing &low psnr value result

[ryujaehun]

examples/super_resolution/data.py

Line 41 in dcdabc2

Resize(crop_size // upscale_factor),

I think resizing LANCZOS interpolation is better than default BILINEAR
Resize(crop_size // upscale_factor,interpolation=Image.LANCZOS)
How does downsampling work in a normal SR?

And In the Set5 dataset, I found that the psnr value is lower than the bicubic method.
Why..?

[chroneus]

yes, LANCZOS is better and PSNR result wich
For data augmentation which aims to different scenarios, even better to use random choice of resize method.

[ryujaehun]

I did downsampling using the LANCZOS method. But, in ESPCN,They used the Gaussian blur method. I can not be sure this is the reason, but the PSNR value is too low. In the Set5 dataset, I found that the psnr value is lower than the bicubic method.

[zpkosmos]

Excuse me, have you solve the problem of low psnr value result ? I tried to transform the 91 images training set into hdf5 file,and trained on it ,but still had 4 dB psnr loss than bicubic method, could you please give me some advises?

[ryujaehun]

Have you measured the PSNR of the entire RGB channel?

[zpkosmos]

Yes, but it still less than the bicubic method, I don't know what is the problem that make this method works worse than bicubic method, I can't sure whether the problem is in training process, SR process or the model.

[ryujaehun]

If so, just use the y channel to see the results.

[zpkosmos]

I mean that I have tried to compute the PSNR on both three channels and y channel, the y channels didn't perform better than it in bicubic, so that it couldn't perform better on three channels. By the way, have you made this project perform better than the bicubic method? What the number of epochs did you set for training to make it perform better than the bicubic?

[Dae12-Han]

I also have the same result. I got the espcn psnr value is lower than the bicubic method. Is there any reason why this pytorch/examples espcn code has lower psnr???