Research map / Image generation / Image restoration
Contrastive Learning for Low-Light Raw Denoising (Student Abstract)
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 denoising contrastive regularization (DCR) with a wavelet-based feature network (Wnet) to improve low-light raw image/video denoising on the starlight dataset of Monakhova et al.
Problem
Denoising raw images and videos captured in extremely low-light conditions, where limited photon counts and complex sensor noise (quantization, bias, clipping) make denoising very hard.
Approach
Applies contrastive learning to low-level denoising: the denoised output is pulled toward the clean image and pushed away from the noisy input in the feature space of a task-specific Wnet (Haar wavelet high-frequency features + CNN, pretrained as a noise classifier on synthetic noisy data), replacing a generic pretrained VGG embedding.
Key results
- On still images in Table 2, Baseline+DCR reaches 30.5 dB PSNR, 0.889 SSIM, 0.060 LPIPS, versus 28.1 dB for Baseline+(VGG+L1) and 27.7 dB for the prior SOTA baseline.
- Training FastDVDnet with DCR improves PSNR from 31.44 to 32.23 dB on the Table 1 still-image test set, while VGG+L1 drops it to 30.23 dB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add DCR | FastDVDnet · Still images test set | PSNR: 31.44 -> 32.23 | Helped |
| Add DCR | FastDVDnet · Still images test set | SSIM: 0.852 -> 0.853 | Helped |
| Add DCR | FastDVDnet · Still images test set | LPIPS: 0.0696 -> 0.0630 | Helped |
| Add DCR | Modified FastDVDnet · Still images test set | PSNR: 28.1 -> 30.5 | Helped |
| Add DCR | Modified FastDVDnet · Still images test set | SSIM: 0.931 -> 0.889 | Hurt |
| Add DCR | Modified FastDVDnet · Still images test set | LPIPS: 0.078 -> 0.060 | Helped |
Surprising findings
- The VGG+L1 contrastive variant lowers PSNR below the baseline (31.44 to 30.23 dB) in Table 1, suggesting generic pretrained features can hurt.
Limitations
- Evaluated only on a single real low-light dataset from Monakhova et al. 2022.
- Wnet pretraining and training details (sizes, hyperparameters, layer weights) are not fully specified in the short paper.
Experimental setup
Datasets: Monakhova et al. 2022 low-light raw image/video dataset (starlight and 0.6 millilux) Synthetic noisy samples from Monakhova et al. 2022 noise generator
Benchmarks: Monakhova et al. 2022 low-light still image and video test set
Metrics: PSNR SSIM LPIPS
Baselines: FastDVDnet Baseline+(VGG+L1) Baseline (SOTA) from Monakhova et al. 2022
Builds on: Monakhova et al. 2022 (starlight dataset and modified FastDVDnet) Contrastive learning for single image dehazing (Wu et al. 2021) FastDVDnet
Closest papers by meaning
- Learning to See in the Dark (CVPR 2018 · similarity 0.89)
- Interpretable Unsupervised Joint Denoising and Enhancement for Real-World low-light Scenarios (ICLR 2025 · similarity 0.87)
- Zero-Shot Low-Light Image Enhancement via Latent Diffusion Models (AAAI 2025 · similarity 0.87)
- DPLUT: Unsupervised Low-light Image Enhancement with Lookup Tables and Diffusion Priors (AAAI 2025 · similarity 0.86)
- DarkIR: Robust Low-Light Image Restoration (CVPR 2025 · similarity 0.86)
- Classic Video Denoising in a Machine Learning World: Robust, Fast, and Controllable (CVPR 2025 · similarity 0.85)
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.