Reproduce “Next-Scale Prediction: A Self-Supervised Approach for Real-World Image Denoising” on OpenResearch

About this paper

Yiwen Shan, Haiyu Zhao, Peng Hu, Xi Peng, Yuanbiao Gou

Next-Scale Prediction (NSP) presents a self-supervised paradigm for real-world image denoising that disentangles noise decorrelation from detail preservation by training a Blind-Spot Network to predict higher-resolution, detailed image patches from lower-resolution, noise-decorrelated inputs. This approach achieves state-of-the-art performance among self-supervised methods on SIDD and DND datasets, also enabling competitive noisy image super-resolution without additional training.

Key insights

• Explicitly decoupling noise decorrelation and detail preservation into separate, scale-dependent operations effectively resolves the inherent trade-off faced by single-scale PD-based BSNs.
• By predicting a higher-resolution, detail-rich target from a lower-resolution, noise-decorrelated input, the model learns to effectively remove noise without sacrificing fine image structures.
• The next-scale prediction mechanism inherently supports noisy image super-resolution as a by-product, providing joint denoising and upsampling capabilities without additional training or model modifications.

Summary by alphaXiv.