Reproduce “Learning Camera-Aware Noise Models” 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
Ke-Chi Chang, Ren Wang, Hung-Jin Lin, Yu-Lun Liu, Chia-Ping Chen, Yu-Lin Chang, Hwann-Tzong Chen
Researchers from MediaTek Inc. and National Tsing Hua University developed a camera-aware, signal-dependent generative model for imaging sensor noise using a GAN framework. This model, trained on raw image data, achieves a D_KL of 0.00159 for noise modeling and enables denoisers to reach 48.72 PSNR and 0.994 SSIM on the SIDD dataset, establishing new state-of-the-art performance.
Key insights
- Combining a statistical prior (Poisson-Gaussian noise) with a deep learning generative model significantly enhances the realism and accuracy of learned noise distributions.
- A dedicated Camera-Encoding Network, regularized by triplet loss, effectively learns and distinguishes unique noise characteristics across diverse camera sensors within a single model.
- A feature matching loss, which aligns discriminator features of generated noise with clean images, is crucial for producing perceptually realistic, signal-dependent noise patterns that mimic real camera artifacts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.