Research map / Image generation / Image restoration
IniRetinex: Rethinking Retinex-type Low-Light Image Enhancer via Initialization Perspective
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 IniRetinex, an unsupervised Retinex-based low-light enhancer that uses a tiny two-layer CNN to compensate the initial illumination and an ADMM/FFT-solved RTV-style optimization with a proximal term to avoid gamma post-processing.
Problem
Low-light image enhancement with Retinex methods suffers from inaccurate initial illumination estimation, which forces gamma-correction post-processing and leads to instability, redundancy, and poor robustness of deep Retinex pipelines.
Approach
Replace the fixed max-RGB initial illumination with a tiny CNN (deep-image-prior style, no pre-training, per-image unsupervised fitting for a few iterations) that adaptively compensates the initial illumination, and optimize smooth illumination via ADMM with an added proximal gradient-fidelity term to mitigate RTV staircase effects.
Key results
- Achieves the lowest NIQE on LIME (3.622), DICM (3.441), NPE (3.583) and Exdark (3.615) among compared methods, per Table 1.
- On LOL, SSIM 0.797 and PSNR 18.247 versus URetinex's 0.822 / 20.141 (URetinex is higher on reference metrics for LOL).
- Uses about 0.0003 s inference per 1080p image on GPU-listed table entries with a tiny parameter count, faster than most deep baselines but slower than SCI and PairLIE due to per-image training.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use IniRetinex instead of competing enhancement methods | Low-light image enhancement · LIME, DICM, NPE, Exdark | NIQE: 3.622; 3.441; 3.583; 3.615 | Helped |
| use IniRetinex instead of competing enhancement methods | Low-light image enhancement · LIME, DICM, NPE, Exdark | IE: 7.433; 7.356; 7.513; 7.518 | Mixed |
| use IniRetinex instead of the strongest listed baselines | Low-light image enhancement · LOL | SSIM: 0.822 -> 0.797 | Hurt |
| use IniRetinex instead of the strongest listed baseline | Low-light image enhancement · LOL | PSNR: 23.081 -> 18.247 | Hurt |
| use IniRetinex instead of the strongest listed baselines | Low-light image enhancement · BAID | SSIM: 0.854 -> 0.793 | Hurt |
| use IniRetinex instead of the strongest listed baselines | Low-light image enhancement · BAID | PSNR: 20.511 -> 17.205 | Hurt |
| train G with 8x downsampled inputs rather than full-resolution inputs (ablation) | IniRetinex adaptive initialization · Ablation study | inference time: 0.391 -> 0.076 | Helped |
| test G with 4x rather than 2x downsampling (ablation) | IniRetinex adaptive initialization · Ablation study | NIQE: 3.491 -> 3.527 | Hurt |
| increase adaptive initialization training iterations (ablation) | IniRetinex network G · 100 randomly selected images | performance: 40 | Mixed |
| add the proximal gradient-preservation term (η) (ablation) | IniRetinex illumination optimization · Low-light image enhancement | illumination edge preservation | Helped |
| increase illumination optimization iterations (ablation) | IniRetinex illumination optimization · Low-light image enhancement | illumination smoothing: 7 | Helped |
Surprising findings
- The method does not achieve the best reference-based scores (e.g., LOL PSNR/SSIM), which the authors attribute to learned baselines being tuned to LOL distribution.
Limitations
- Per-image optimization makes inference slower than SCI and PairLIE.
- Reference-based metrics (SSIM/PSNR/LPIPS) on LOL trail some learned baselines such as URetinex.
- Relies on a pre-trained DANet for denoising in noisy real-world scenarios.
Experimental setup
Datasets: LIME DICM NPE Exdark LOL BAID
Benchmarks: LIME DICM NPE Exdark LOL BAID
Metrics: NIQE IE SSIM PSNR LPIPS inference time parameter count
Baselines: LIME RRM Zero-DCE++ SCI URetinex PairLIE NeRCo ZRLLE-PQP
Builds on: Retinex theory RTV (relative total variation) deep image prior LIME Zero-DCE++
Closest papers by meaning
- MambaLLIE: Implicit Retinex-Aware Low Light Enhancement with Global-then-Local State Space (NeurIPS 2024 · similarity 0.89)
- Enhancing Low-Light Images: A Synthetic Data Perspective on Practical and Generalizable Solutions (AAAI 2025 · similarity 0.88)
- Zero-Shot Low-Light Image Enhancement via Latent Diffusion Models (AAAI 2025 · similarity 0.87)
- Dehaze-RetinexGAN: Real-World Image Dehazing via Retinex-based Generative Adversarial Network (AAAI 2025 · similarity 0.86)
- Reti-Diff: Illumination Degradation Image Restoration with Retinex-based Latent Diffusion Model (ICLR 2025 · similarity 0.86)
- Efficient Diffusion as Low Light Enhancer (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.