PSNR loss and L1 Loss - Githubissues

megvii-research / NAFNet

The state-of-the-art image restoration model without nonlinear activation functions.

Other

2.21k stars 270 forks source link

PSNR loss and L1 Loss #50

Open dasonglicu opened 2 years ago

dasonglicu commented 2 years ago

I want to know the quantitative comparisons between PSNR loss and L1 loss.

Since PSNR loss is adopted in NAFNet and HINet, I guess the PSNR loss achieves the better performance.

By the way, why PSNR loss is better? any gradient or optimization analysis?

mayorx commented 2 years ago

Hi, dmeatfloss,

When we participated in the NTIRE 2021, deblur competition, we suddenly thought "why not directly use the differentiable metric of PSNR as the optimization target, but use the indirect L1/L2 losses?". So we tried PSNR loss and our ablation experiments show that it is slightly better than L1/L2 (+0.01~0.02 dB) at that time. So it has been used until today.