File(s) under permanent embargo
Lightweight Wavelet-Based Transformer for Image Super-Resolution
Suffering from the inefficiency of deeper and wider networks, most remarkable super-resolution algorithms cannot be easily applied to real-world scenarios, especially resource-constrained devices. In this paper, to concentrate on fewer parameters and faster inference, an end-to-end Wavelet-based Transformer for Image Super-resolution (WTSR) is proposed. Different from the existing approaches that directly map low-resolution (LR) images to high-resolution (HR) images, WTSR also implicitly mines the self-similarity of image patches by a lightweight Transformer on the wavelet domain, so as to balance the model performance and computational cost. More specifically, a two-dimensional stationary wavelet transform is designed for the mutual transformation between feature maps and wavelet coefficients, which reduces the difficulty of mining self-similarity. For the wavelet coefficients, a Lightweight Transformer Backbone (LTB) and a Wavelet Coefficient Enhancement Backbone (WECB) are proposed to capture and model the long-term dependency between image patches. Furthermore, a Similarity Matching Block (SMB) is investigated to combine global self-similarity and local self-similarity in LTB. Experimental results show that our proposed approach can achieve better super-resolution performance on the multiple public benchmarks with less computational complexity.
History
Volume
13631 LNCSPagination
368-382Publisher DOI
ISSN
0302-9743eISSN
1611-3349ISBN-13
9783031208676Publication classification
E1.1 Full written paper - refereedTitle of proceedings
Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)Publisher
Springer Nature SwitzerlandSeries
Lecture Notes in Computer ScienceUsage metrics
Categories
No categories selectedKeywords
Licence
Exports
RefWorks
BibTeX
Ref. manager
Endnote
DataCite
NLM
DC