Reproduce “Estimating Fine-Grained Noise Model via Contrastive Learning” 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
Yunhao Zou, Ying Fu
Researchers at Beijing Institute of Technology developed a method for estimating and synthesizing realistic camera noise using a fine-grained physics-based model and contrastive learning. This approach accurately models camera-specific noise without requiring calibration data or paired images, leading to synthetic training data that enables denoisers to perform comparably to those trained on real data.
Key insights
- A fine-grained, physics-based noise model incorporating shot noise, readout noise, color bias, and row noise accurately represents real-world sensor noise.
- A consistent logarithmic linear relationship exists between camera gain (K) and other noise parameters (readout noise, row noise) across different camera models, which can be leveraged for accurate parameter sampling.
- Contrastive learning enables the estimation of camera-specific noise parameters without requiring paired clean/noisy images or camera calibration frames.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.