By Jiahui Yu1, Yuchen Fan1, Jianchao Yang2, Ning Xu2, Zhaowen Wang3, Xinchao Wang4, Thomas S. Huang1
[1] University of Illinois at Urbana-Champaign, [2] Snap Inc., [3] Adobe Research, [4] Stevens Institute of Technology| Network | Parameters | DIV2K (val) PSNR |
|---|---|---|
| EDSR Baseline | 1,372,318 | 34.61 |
| WDSR Baseline | 1,190,100 | 34.77 |
We measured PSNR using DIV2K 0801 ~ 0900 (trained on 0000 ~ 0800) on RGB channels without self-ensemble which is identical to EDSR baseline model settings. Both baseline models have 16 residual blocks.
More results:
| Number of Residual Blocks | 1 | 3 | ||
|---|---|---|---|---|
| SR Network | EDSR | WDSR | EDSR | WDSR |
| Parameters | 2.6M | 0.8M | 4.1M | 2.3M |
| DIV2K (val) PSNR | 33.210 | 33.323 | 34.043 | 34.163 |
| Number of Residual Blocks | 5 | 8 | ||
|---|---|---|---|---|
| SR Network | EDSR | WDSR | EDSR | WDSR |
| Parameters | 5.6M | 3.7M | 7.8M | 6.0M |
| DIV2K (val) PSNR | 34.284 | 34.388 | 34.457 | 34.541 |
Comparisons of EDSR and our proposed WDSR for image bicubic x2 super-resolution on DIV2K dataset.
Left: vanilla residual block in EDSR. Middle: wide activation. Right: wider activation with factorized convolution. The proposed wide activation has similar merits with MobileNet V2.
Training loss and validation PSNR with weight normalization, batch normalization or no normalization. Training with weight normalization has better performance.