the psnr becomes terrible after adding jpeg compression noise. re-training with the noised data does not help. is there a cure?

JiahuiYu / wdsr_ntire2018

Code of our winning entry to NTIRE super-resolution challenge, CVPR 2018

http://www.vision.ee.ethz.ch/ntire18/

597 stars 123 forks source link

the psnr becomes terrible after adding jpeg compression noise. re-training with the noised data does not help. is there a cure? #22

Closed JVision closed 5 years ago

JVision commented 5 years ago

Thanks for the great work. Really inspiring. Though, an issue was found when I am playing with the real data. for images destroyed by jpeg compression, the improvement is usually minor. I tried re-train the neural nets with jpeg images. However, it does not help much, usually giving a psnr value close to 31.7db, significantly lower than its bicubic part. Is there any solution? Thanks in advance.

JiahuiYu commented 5 years ago

JPEG compression artifacts are more difficult to restore, thus lower PSNR does not surprise me. I would suggest you to try other models and compare with JPEG restoration baselines, instead of comparing with bicubic one.

JVision commented 5 years ago

thanks for the prompt answer. Is there any solution to this problem? In real use case, most images are simply compressed. Really appreciated if any insights could be given.

JiahuiYu commented 5 years ago

I would suggest you to read related papers on JPEG artifacts removal. Also, @ychfan had a paper on how to do restoration for BPG-compressed images based on WDSR, which won honorable mention award and ranked 4th in image compression challenge in CVPR 2018.

Wide-activated Deep Residual Networks based Restoration for BPG-compressed Images