Reproduce “Image Restoration via Multi-domain Learning” 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
Xingyu Jiang, Ning Gao, Xiuhui Zhang, Hongkun Dou, Shaowen Fu, Xiaoqing Zhong, Hongjue Li, Yue Deng
The paper introduces SWFormer, an image restoration framework leveraging multi-domain learning across spatial, wavelet, and Fourier domains within an efficient Transformer-like architecture. It achieves state-of-the-art performance across 10 diverse image restoration tasks while significantly reducing model complexity and inference latency, demonstrating up to 67 times faster processing than prior methods.
Key insights
- Degradations, while diffuse spatially, exhibit compact and localized patterns in the Fourier domain, and specific frequency bands are affected differently by various degradation types.
- Wavelet Transform can powerfully separate image components into sub-bands, revealing that degradations concentrate in specific sub-bands and frequency ranges when combined with Fourier analysis.
- Efficient multi-domain modeling can be achieved by leveraging the spatial domain for local features, the wavelet domain for regional features, and the Fourier domain for global dependencies, where its O(N log N) complexity is superior to O(N^2) self-attention.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.