Research map / Image generation / Image restoration
HVI: A New Color Space for Low-light Image Enhancement
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 the HVI color space (polarized HS maps plus a learnable intensity collapse function) and CIDNet, a dual-branch UNet with cross-attention, for low-light image enhancement.
Problem
Low-light image enhancement in sRGB causes color bias and brightness artifacts, while HSV conversion introduces red-discontinuity and black-plane noise artifacts.
Approach
Defines a new HVI color space by polarizing the HS plane (cos/sin of hue) and scaling it with a trainable intensity collapse function Ck so that similar colors (especially reds) and dark pixels have small Euclidean distances; a dual-branch network (HV-branch and I-branch with cross-attention) then enhances the decoupled components.
Key results
- CIDNet reaches 28.201 PSNR / 0.889 SSIM on LOLv1 and 24.111 PSNR on LOLv2-Real with 1.88M params and 7.57 GFLOPs.
- On Sony-Total-Dark, CIDNet surpasses the second-best method by 6.678 dB PSNR (22.904 vs. 16.226 for LLFlow).
- Plugging HVI transform into GSAD raises its LOLv2-Real PSNR by 3.562 dB (to 23.715).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace sRGB with HVI color space (ablation) | CIDNet · LOLv2-Real | PSNR: 20.062 -> 24.111 | Helped |
| add HS polarization to HSV (ablation) | CIDNet color representation · LOLv2-Real | PSNR: 21.349 -> 21.558 | Helped |
| add the Ck intensity collapse function to HSV (ablation) | CIDNet color representation · LOLv2-Real | PSNR: 21.349 -> 21.536 | Helped |
| combine HS polarization and Ck intensity collapse (ablation) | CIDNet color representation · LOLv2-Real | PSNR: 21.349 -> 24.111 | Helped |
| add self-attention (ablation) | UNet baseline · LOLv2-Real | PSNR: 19.306 -> 22.313 | Helped |
| add a dual-branch structure without cross-attention (ablation) | SelfAttn network · LOLv2-Real | PSNR: 22.313 -> 23.159 | Helped |
| replace self-attention with cross-attention between dual branches (ablation) | Dual-branch network · LOLv2-Real | PSNR: 23.159 -> 24.111 | Helped |
| use both HVI and sRGB losses instead of HVI loss only (ablation) | HVI-CIDNet · LOLv2-Real | PSNR: 23.221 -> 24.111 | Helped |
| use both HVI and sRGB losses instead of sRGB loss only (ablation) | HVI-CIDNet · LOLv2-Real | PSNR: 23.319 -> 24.111 | Helped |
| apply HVI transformation as a plug-in | FourLLIE · LOLv2-Real | PSNR: +0.381 | Helped |
| apply HVI transformation as a plug-in | LEDNet · LOLv2-Real | PSNR: +3.456 | Helped |
| apply HVI transformation as a plug-in | SNR-Aware · LOLv2-Real | SSIM: -0.009 | Hurt |
Surprising findings
- Using only the polarization or only the Ck intensity function in HVI yields PSNR similar to HSV, and Ck alone causes red/color confusion artifacts.
- The HSV color space improves PSNR over sRGB but degrades SSIM and LPIPS due to red-discontinuity noise.
Limitations
- Reliance on GT-mean testing for LOLv1 due to a small test set, which the authors note as a methodological workaround.
- The HVI color space is built on HSV and inherits some of its dependence on the Max-RGB intensity estimate.
Experimental setup
Datasets: LOLv1 LOLv2 SICE Sony-Total-Dark DICM LIME MEF NPE VV
Benchmarks: LOLv1 LOLv2-Real LOLv2-Synthetic SICE Sony-Total-Dark DICM LIME MEF NPE VV
Metrics: PSNR SSIM LPIPS BRISQUE NIQE
Baselines: RetinexNet KinD ZeroDCE RUAS LLFlow EnlightenGAN SNR-Aware Bread PairLIE LLFormer RetinexFormer GSAD
Builds on: HSV color space Max-RGB RetinexFormer UNet
Closest papers by meaning
- Enhancing Low-Light Images: A Synthetic Data Perspective on Practical and Generalizable Solutions (AAAI 2025 · similarity 0.86)
- IniRetinex: Rethinking Retinex-type Low-Light Image Enhancer via Initialization Perspective (AAAI 2025 · similarity 0.84)
- Zero-Shot Low-Light Image Enhancement via Latent Diffusion Models (AAAI 2025 · similarity 0.83)
- Luminance-Aware Statistical Quantization: Unsupervised Hierarchical Learning for Illumination Enhancement (NeurIPS 2025 · similarity 0.83)
- Efficient Diffusion as Low Light Enhancer (CVPR 2025 · similarity 0.82)
- Color Shift Estimation-and-Correction for Image Enhancement (CVPR 2024 · similarity 0.82)
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.