Reproduce “Dual-Camera Super-Resolution with Aligned Attention Modules” 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
Tengfei Wang, Jiaxin Xie, Wenxiu Sun, Qiong Yan, Qifeng Chen
The paper introduces a learning-based framework for real-world dual-camera super-resolution (DCSR) that integrates an aligned attention module for robust feature warping and an adaptive fusion module for detail transfer. It achieved superior perceptual quality and quantitative metrics on both public benchmarks and a new 4K real-world smartphone dataset.
Key insights
- Explicitly learning local transformations for individual matched patches can robustly handle complex spatial misalignments in reference-based super-resolution.
- Incorporating learnable confidence maps for adaptive fusion in both feature and image spaces allows for intelligent integration of reference details, leading to sharper reconstructions.
- A targeted fidelity loss is crucial to ensure the network effectively utilizes the high-resolution reference, compelling it to transfer detailed information.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.