Research map / Image generation / Image restoration
Compress Clean Signal from Noisy Raw Image: A Self-Supervised Approach
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a self-supervised joint denoising and compression scheme for raw images that compresses only the clean signal using a physics-based noise model, and introduces the FDRIC full-day raw dataset.
Problem
Raw images carry strong sensor noise that inflates bitrate and lowers reconstruction quality in learned raw image compression, and existing noise-aware methods need clean paired data that is impractical to collect for raw images.
Approach
Jointly denoise and compress raw images by having a learned noise extractor regularized by a physics-based noise model (heteroscedastic Gaussian, row noise, fixed pattern noise) via normalizing-flow-style bijective transforms and an NLL plus covariance loss, so that noise is excluded from the bitstream without clean ground truth.
Key results
- Compared to RIC on SID, achieves roughly tenfold bitrate reduction and more than 8 dB PSNR gain.
- BD-rate of 0% for the proposed method vs 234.55% for BM3D+RIC and 317.69% for Ne2Ne+RIC on FDRIC, with about 500x faster encoding than BM3D+RIC.
- Outperforms supervised SID+RIC (8.35% BD-rate) and ELD+RIC (59.50%) on SID without any paired clean data.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| selectively compress the clean signal while disentangling and discarding noise | Ours vs. RIC · SID SonyA7S2 sample | PSNR (dB): 41.03 -> 41.17 | Helped |
| selectively compress the clean signal while disentangling and discarding noise | Ours vs. RIC · FDRIC OV64B sample | PSNR (dB): 44.89 -> 49.14 | Helped |
| use the self-supervised method instead of SID+RIC | SID, ratio=100 · SID | BD-Rate↓(%): 8.35 -> 0 | Helped |
| use the self-supervised method instead of ELD+RIC | SID, ratio=100 · SID | BD-Rate↓(%): 59.50 -> 0 | Helped |
| replace BM3D+RIC with the proposed method | FDRIC · FDRIC | Latency (s): 290.629 -> 0.567 | Helped |
| replace BM3D+RIC with the proposed method | FDRIC · FDRIC | BD-rate (%): 234.55 -> 0 | Helped |
| replace Ne2Ne+RIC with the proposed method | FDRIC · FDRIC | BD-rate (%): 317.69 -> 0 | Helped |
| reduce λP to 0.1 (ablation) | Ours · FDRIC | BD-Rate↓(%): 0 -> 48.88 | Hurt |
| increase λP to 30 (ablation) | Ours · FDRIC | BD-Rate↓(%): 0 -> 28.64 | Hurt |
| set λcov to 0.0 (ablation) | Ours · FDRIC | BD-Rate↓(%): 0 -> 5.35 | Hurt |
| set λcov to 0.02 (ablation) | Ours · FDRIC | BD-Rate↓(%): 0 -> 4.45 | Hurt |
| set λcov to 2 (ablation) | Ours · FDRIC | BD-Rate↓(%): 0 -> 3.80 | Hurt |
Surprising findings
- Self-supervised method outperforms supervised two-stage denoiser+compressor pipelines trained on paired SID data.
Limitations
- Requires calibrated camera noise model parameters per camera
- Evaluated on a limited set of cameras and 32 test pairs in FDRIC
Experimental setup
Datasets: SID ELD FDRIC
Benchmarks: SID ELD FDRIC
Metrics: PSNR bpp BD-rate FLOPs latency
Baselines: RIC BM3D+RIC Neighbor2Neighbor+RIC LGBPN+RIC SID+RIC ELD+RIC QARV
Builds on: RIC Ballé hyperprior compression ELD noise model PMN Cheng et al. 2022 joint denoising and compression QARV
Closest papers by meaning
- Towards image compression with perfect realism at ultra-low bitrates (ICLR 2024 · similarity 0.88)
- Noise Modeling in One Hour: Minimizing Preparation Efforts for Self-supervised Low-Light RAW Image Denoising (CVPR 2025 · similarity 0.87)
- C3: High-Performance and Low-Complexity Neural Compression from a Single Image or Video (CVPR 2024 · similarity 0.86)
- Good, Cheap, and Fast: Overfitted Image Compression with Wasserstein Distortion (CVPR 2025 · similarity 0.86)
- Make Lossy Compression Meaningful for Low-Light Images (AAAI 2024 · similarity 0.86)
- Compressed Image Generation with Denoising Diffusion Codebook Models (ICML 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.