Research map / Image generation / Image restoration
Interpretable Unsupervised Joint Denoising and Enhancement for Real-World low-light Scenarios
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 zero-reference, interpretable framework for joint low-light enhancement and denoising that builds paired sub-images via neighboring-pixel masking and random gamma correction, guided by Retinex decomposition and DCT-based frequency-illumination priors.
Problem
Real-world low-light images suffer compounded degradations (overexposure, low brightness, noise, uneven illumination), and supervised methods overfit while unsupervised/zero-reference methods struggle to model these degradations without reference images.
Approach
Generates two sub-images of the same scene via 2x2 neighboring-pixel masking, applies gamma correction to one to create an illumination difference, and enforces reflectance consistency under Retinex decomposition, with a DCT-based frequency-illumination prior encoder providing implicit degradation representations via cross-attention into the reflectance network.
Key results
- On LOLv2-Real, the method reports PSNR 20.22, SSIM 0.793 and LPIPS 0.266, the best among compared methods in the paper's table.
- On SICE, it reports PSNR 22.55 and SSIM 0.841, with 0.36M parameters, outperforming supervised and unpaired baselines in the table.
- On SIDD, it achieves the best BRISQUE (2.555) and CLIPIQA (0.292) among the compared methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the full DEnet method rather than SNR-aware | low-light enhancement · LOLv1 | PSNR: 24.61 -> 19.80 | Hurt |
| add the full DEnet method rather than LLFormer | low-light enhancement · LOLv2-Real | PSNR: 27.75 -> 20.22 | Hurt |
| add the full DEnet method rather than PairLIE | low-light enhancement · SICE | PSNR: 21.32 -> 22.55 | Helped |
| add the full DEnet method rather than PairLIE | low-light enhancement · SIDD | BRISQUE: 3.168 -> 2.555 | Helped |
| add illumination, low-pass, and high-pass physical priors (ablation) | DEnet · LOLv1 | PSNR: 18.88 -> 19.80 | Helped |
| add illumination, low-pass, and high-pass physical priors (ablation) | DEnet · LOLv2 | PSNR: 19.37 -> 20.22 | Helped |
| add low-pass prior to illumination-only priors (ablation) | DEnet · LOLv1 | PSNR: 19.54 -> 19.57 | Helped |
| add low-pass prior to illumination-only priors (ablation) | DEnet · LOLv2 | PSNR: 19.99 -> 19.51 | Hurt |
| add high-pass prior to illumination-only priors (ablation) | DEnet · LOLv1 | PSNR: 19.54 -> 19.69 | Helped |
| add high-pass prior to illumination-only priors (ablation) | DEnet · LOLv2 | PSNR: 19.99 -> 19.28 | Hurt |
| add neighborhood masking (ablation) | DEnet · LOLv1 | PSNR: 18.52 -> 19.63 | Helped |
| add neighborhood masking (ablation) | DEnet · LOLv2 | PSNR: 19.46 -> 19.83 | Helped |
Surprising findings
- Removing the neighboring-pixel masking and training on full-resolution images with different illumination causes the decomposition network to learn near-identity mappings, increasing noise.
Limitations
- Gamma factor must be near 1 for the Taylor-approximation assumption to hold; larger factors degrade performance.
- Relies on hand-set DCT bandwidth hyperparameter t and multiple loss weights.
- Evaluation on SIDD relies on no-reference metrics only, with no full-reference comparison.
Experimental setup
Datasets: LOLv1 LOLv2-Real SICE SIDD
Benchmarks: LOLv1 LOLv2-Real SICE SIDD
Metrics: PSNR SSIM LPIPS BRISQUE CLIPIQA
Baselines: URetinexNet SNR-aware LLFormer Retinexformer Retinexmamba EnlightenGAN PairLIE NeRCo Zero-DCE RUAS SCI Clip-LIT
Builds on: Retinex theory Noise2Noise Zero-DCE PairLIE
Closest papers by meaning
- ZERO-IG: Zero-Shot Illumination-Guided Joint Denoising and Adaptive Enhancement for Low-Light Images (CVPR 2024 · similarity 0.91)
- AGLLDiff: Guiding Diffusion Models Towards Unsupervised Training-free Real-world Low-light Image Enhancement (AAAI 2025 · similarity 0.89)
- Zero-Shot Low-Light Image Enhancement via Latent Diffusion Models (AAAI 2025 · similarity 0.88)
- Zero-Reference Low-Light Enhancement via Physical Quadruple Priors (CVPR 2024 · similarity 0.88)
- Fourier Priors-Guided Diffusion for Zero-Shot Joint Low-Light Enhancement and Deblurring (CVPR 2024 · similarity 0.87)
- Enhancing Low-Light Images: A Synthetic Data Perspective on Practical and Generalizable Solutions (AAAI 2025 · similarity 0.87)
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.