Abstract
In this work we demonstrate that with same parameters and computational budgets, models with wider features before ReLU activation have significantly better performance for image and video super-resolution. The resulted SR residual network has a slim identity mapping pathway with wider (2× to 4×) channels before activation in each residual block. To further widen activation (6× to 9×) without computational overhead, we introduce linear low-rank convolution into SR networks and achieve even better accuracy-efficiency tradeoffs. In addition, compared with batch normalization or no normalization, we find training with weight normalization leads to better accuracy for deep super-resolution networks. Our proposed SR network WDSR achieves better results on large-scale DIV2K image super-resolution benchmark in terms of PSNR, under same or lower computational complexity. Based on WDSR, our method won 1st places in NTIRE 2018 Challenge on Single Image Super-Resolution in all three realistic tracks. Moreover, a simple frame-concatenation based WDSR achieved 2nd places in three out of four tracks of NTIRE 2019 Challenge for Video Super-Resolution and Deblurring. Our experiments and ablation studies support the importance of wide activation. Code and models will be publicly available.
Original language | English (US) |
---|---|
State | Published - 2020 |
Event | 30th British Machine Vision Conference, BMVC 2019 - Cardiff, United Kingdom Duration: Sep 9 2019 → Sep 12 2019 |
Conference
Conference | 30th British Machine Vision Conference, BMVC 2019 |
---|---|
Country/Territory | United Kingdom |
City | Cardiff |
Period | 9/9/19 → 9/12/19 |
ASJC Scopus subject areas
- Computer Vision and Pattern Recognition