Reproduce “BSRT: Improving Burst Super-Resolution with Swin Transformer and Flow-Guided Deformable Alignment” 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
Ziwei Luo, Youwei Li, Shen Cheng, Lei Yu, Qi Wu, Zhihong Wen, Haoqiang Fan, Jian Sun and 1 more
The BSRT framework introduces a Burst Super-Resolution Transformer that reconstructs high-quality images from noisy, low-resolution RAW burst photos by integrating a Swin Transformer backbone with a Pyramid Flow-Guided Deformable Alignment network. This approach achieved top performance, winning the NTIRE2022 Burst Super-Resolution Challenge in the real-world track.
Key insights
- Vision Transformers, specifically the Swin Transformer's shifted window mechanism, are effective in capturing global contextual information and long-range dependencies crucial for superior information aggregation in burst super-resolution.
- A hybrid alignment strategy that integrates coarse optical flow guidance with fine, learned deformable convolution offsets within a multi-scale pyramid structure robustly handles complex inter-frame misalignments and noise.
- Optimizing the RAW image processing workflow by extracting initial features at the original low resolution before upscaling for alignment significantly reduces computational cost while preserving or improving reconstruction quality.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.