Efficient Single Image Super-Resolution Using Compact Back-Projection Network and Knowledge Distillation
Deep learning methods have achieved state-of-the-art accuracy in single image super-resolution (SISR). Yet, how to achieve good balance between efficiency and accuracy in SISR is still an open issue. While most existing methods learn residual features only in low-resolution (LR) space in order for high efficiency, recent studies show that jointly learning residual features in LR and high-resolution (HR) spaces is more conducive to accurate SISR. In this paper, we propose an efficient SISR scheme via learning hybrid residual features, based on which the residual HR image is reconstructed. To fulfill hybrid residual feature learning, we propose a novel compact back-projection network (CBPN) that simultaneously generates features in both LR and HR space by cascading up- and down-sampling layers with small-size filters. To further enhance the efficiency and yield a more compact model, we apply knowledge distillation to the proposed CBPN on pixel-level space and obtain improved performance. Our knowledge distillation is inspired by curriculum learning, which encourages learning from simple to difficult. Extensive experiments on four benchmark datasets demonstrate that our proposed CBPN achieves high efficiency (i.e., small number of parameters and operations) meanwhile preserving state-of-the-art SR accuracy
Year of publication: |
[2022]
|
---|---|
Authors: | Zhu, Feiyang ; Fu, Keren ; Zhao, Qijun |
Publisher: |
[S.l.] : SSRN |
Subject: | Theorie | Theory | Wissenstransfer | Knowledge transfer | Unternehmensnetzwerk | Business network | Effizienz | Efficiency | Wissen | Knowledge |
Saved in:
freely available
Saved in favorites
Similar items by subject
-
Overcoming undesirable knowledge redundancy in territorial clusters
Molina-Morales, F. Xavier, (2013)
-
Regional efficiency in generating technological knowledge
Schaffer, Axel, (2010)
-
Böhn, Thorsten, (2006)
- More ...