Reproduce “Patch-Craft Self-Supervised Training for Correlated Image Denoising” 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
Gregory Vaksman, Michael Elad
This paper presents the Patch-Craft (PC) framework for self-supervised image denoising that handles unknown, spatially, and cross-channel correlated noise by generating synthetic training targets from bursts of noisy images. The method achieves superior denoising performance on both synthetic and real-world correlated noise, outperforming existing state-of-the-art self-supervised techniques.
Key insights
- Patch matching across different noisy frames in a burst can generate effective surrogate clean targets for self-supervised denoising, provided the noise realizations are sufficiently independent.
- Statistical dependencies can arise between the noisy input and the patch-crafted target noise, negatively impacting training; analyzing and pruning problematic training pairs based on their noise covariance (Type II dependency) significantly improves denoising performance.
- The patch size used for crafting targets should be large, scaling with noise standard deviation and correlation range, to ensure effective content matching over noise.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.