Reproduce “FourierSR: A Fourier Token-based Plugin for Efficient Image Super-Resolution” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Wenjie Li, Heng Guo, Yuefeng Hou, Zhanyu Ma
FourierSR introduces a lightweight, plug-and-play module that enhances efficient image super-resolution models by leveraging a Fourier token-based mechanism for global receptive fields. It achieves an average PSNR gain of 0.18 dB across benchmarks, with only marginal increases in parameters (0.6%) and FLOPs (2%).
Key insights
- Global receptive fields can be efficiently achieved in image super-resolution by performing token mixing and convolution-like operations in the Fourier domain.
- Modulating the real and imaginary parts of Fourier features separately with flexible, broadcast-expanded local filters is crucial for stable training and adapting to varying input sizes.
- The combination of channel token mixing and distinct processing of real/imaginary Fourier components contributes to robust feature learning and enhanced reconstruction quality.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.