PaperlayerView exact revision on arXiv

Source-linked AI summary

HVI: A New Color Space for Low-light Image Enhancement

Qingsen Yan, Yixu Feng, Cheng Zhang, Guansong Pang, Kangbiao Shi, Peng Wu, Wei Dong, Jinqiu Sun, Yanning Zhang

arXiv:2502.20272v2cs.CVcs.AIcs.LG

TL;DR

LLIE methods face color bias and brightness artifacts because sRGB couples color and brightness, while HSV introduces red and black noise. The paper proposes HVI, combining polarized HS maps with learnable intensity, and CIDNet, which separately models chromatic and intensity information. Across 10 datasets, HVI with CIDNet outperforms state-of-the-art LLIE methods, although broader task and training paradigms remain unexplored.

  • Problem

    sRGB-based LLIE suffers color distortion from brightness–color coupling, while HSV enhancement introduces red discontinuity and black-region artifacts.

  • Method

    HVI uses polarized HS maps and trainable intensity collapse, while CIDNet decouples chromatic and intensity modeling for photometric enhancement.

  • Results

    HVI with CIDNet outperforms state-of-the-art LLIE methods on different metrics across 10 datasets.

  • Takeaways & Limitations

    HVI provides a robust LLIE color space that preserves brightness–color decoupling while addressing red and black noise artifacts.

  • Takeaways & Limitations

    The study does not evaluate HVI on additional vision tasks or unsupervised, semi-supervised, and zero-shot training paradigms.

Abstract

from arXiv · show

Low-Light Image Enhancement (LLIE) is a crucial computer vision task that aims to restore detailed visual information from corrupted low-light images. Many existing LLIE methods are based on standard RGB (sRGB) space, which often produce color bias and brightness artifacts due to inherent high color sensitivity in sRGB. While converting the images using Hue, Saturation and Value (HSV) color space helps resolve the brightness issue, it introduces significant red and black noise artifacts. To address this issue, we propose a new color space for LLIE, namely Horizontal/Vertical-Intensity (HVI), defined by polarized HS maps and learnable intensity. The former enforces small distances for red coordinates to remove the red artifacts, while the latter compresses the low-light regions to remove the black artifacts. To fully leverage the chromatic and intensity information, a novel Color and Intensity Decoupling Network (CIDNet) is further introduced to learn accurate photometric mapping function under different lighting conditions in the HVI space. Comprehensive results from benchmark and ablation experiments show that the proposed HVI color space with CIDNet outperforms the state-of-the-art methods on 10 datasets. The code is available at https://github.com/Fediory/HVI-CIDNet.

1. Introduction

LLIE must recover brightness while reducing noise and color bias, but sRGB and HSV representations introduce distinct artifacts. HVI and CIDNet address these issues by decoupling color and intensity information for enhancement.

  • Low-light enhancement targets brightness improvement while reducing severe sensor noise and color bias.
  • sRGB-based methods couple brightness with color, causing restored-image color distortion, while HSV improves brightness enhancement but amplifies red and black artifacts.
  • HVI polarizes the HS plane to reduce distances between similar red coordinates and uses trainable intensity collapse to compress low-light regions.The intensity function preserves the primary image appearance while removing black noise artifacts.
  • CIDNet separately models decoupled color and brightness information through HV and intensity branches to learn photometric mappings under varying lighting.
  • 1.88M parameters and 7.57GFLOPs support a lightweight CIDNet that outperforms state-of-the-art methods across metrics and 10 datasets.

2. Related Work

Prior LLIE methods use single-stage, Retinex, color-space, and diffusion-based strategies, but reported approaches retain brightness, color, or artifact-related challenges.

  • Single-stage methods include RetinexNet and Bread, which decouple illumination or color components but can retain inaccurate brightness control and biased colors in black areas.
  • Diffusion-based LLIE methods can generate accurate images but still exhibit local overexposure or color shifts.
  • sRGB is widely used in imaging, but brightness and color are strongly interdependent across its three channels.
  • HSV decouples brightness from color, yet red discontinuities and black-plane noise create artifacts during enhancement.
  • YCbCr-based methods address hue-related issues but are also associated with severe color shifts.

3. HVI Color Space

HVI retains HSV’s brightness–color decoupling while addressing red discontinuity and black-plane noise through polarized HS coordinates and adaptive intensity collapse.

  • HVI builds on HSV to preserve brightness–color decoupling while minimizing color-space artifacts.
  • The intensity map is estimated from the sRGB input using Max-RGB theory rather than a neural network-generated normal-light map.
  • HSV improves brightness balance but amplifies noise in red discontinuities and black regions, degrading perceptual quality.
  • Horizontal/Vertical Plane with Polarized HS and Collapsible Intensity: Polarizing the Hue axis maps the hue into an orthogonalized H–V plane, preventing identical red colors at h = 0 and h = 6 from being separated.
  • Horizontal/Vertical Plane with Polarized HS and Collapsible Intensity: The adaptive collapse function Ck learns dark color-point density and maps low-intensity radii smaller as Ck decreases, clustering black points.The trainable parameter k controls density, while ε = 1×10−8 avoids gradient explosion.
  • Horizontal/Vertical Plane with Polarized HS and Collapsible Intensity: The resulting ˆH, ˆV, and Imax maps form HVI, whose trainable components support LLIE optimization under varied lighting conditions.

4. Color and Intensity Decoupling Network

CIDNet transforms sRGB images into HVI, enhances decoupled chromatic and intensity information with separate branches, and maps the result back to sRGB under dual-space supervision.

  • CIDNet uses HV-branch processing for dark-region noise and chromaticity suppression, while I-branch processing estimates whole-image illuminance.
  • The framework applies HVI transformation, dual-branch enhancement, and perceptual-inverse HVI transformation consecutively.
  • HVI transformation produces an intensity map containing illuminance and an HV color map containing color and structure information.
  • The dual-branch UNet uses an encoder and decoder with three LCA modules each, plus skip connections.
  • Separate I- and HV-branches address brightness enhancement and low-light noise removal because the subtasks follow distinct statistical patterns.
  • Cross-attention uses intensity features to guide HV denoising, reducing global color shifts and improving noise suppression.
  • PHVIT maps HVI back toward HSV and permits independent adjustment of saturation and brightness before producing the sRGB image.
  • Training jointly supervises enhanced HVI and restored sRGB outputs against their corresponding targets, balancing the two losses with λ.

5. Experiments

Experiments evaluate CIDNet across paired, unpaired, and extreme low-light datasets, then test HVI plug-in generality and architectural components. CIDNet achieves strong restoration quality while balancing efficiency, and ablations attribute gains to combining HVI polarization, adaptive intensity, dual branches, and cross-attention.

  • Datasets and Settings: CIDNet is evaluated on seven common LLIE benchmarks, two extreme datasets, and five unpaired datasets using distortion and perceptual metrics.Paired datasets use PSNR, SSIM, and LPIPS; unpaired datasets use BRISQUE and NIQE.
  • Main Results: CIDNet achieves optimal PSNR, SSIM, and LPIPS across LOLv1 and LOLv2 with 1.88M parameters and 7.57 GFLOPs.It outperforms GSAD across these metrics while using only 10.8% of GSAD’s parameters.
  • Main Results: 6.678 dB PSNR separates CIDNet from the second-best method on Sony-Total-Dark, where its intensity collapse function supports detail recovery under extreme darkness.CIDNet is also reported as the best performer in PSNR and SSIM on SICE and Sony-Total-Dark.
  • Main Results: HVI improves PSNR, SSIM, and LPIPS when applied as a plug-in to multiple sRGB-based LLIE methods, with GSAD gaining 3.562 dB in PSNR.The results support HVI’s generalizability across methods with differing architectures.
  • Main Results: CIDNet provides the most efficient inference while achieving the highest PSNR and second-best SSIM and LPIPS, contrasting with slower diffusion-based methods.Diffusion methods require longer GPU time but achieve better enhancement results in the reported comparison.
  • Ablation Study: Combining polarization and the adaptive intensity function mitigates HSV red discontinuity and black artifacts, producing consistent improvement across PSNR, SSIM, and LPIPS.The full model also benefits from dual branches and cross-attention, while the dual-branch change alone raises PSNR by 0.846 dB.

6. Conclusion

HVI and CIDNet address color bias and brightness artifacts in sRGB-based LLIE by decoupling chromatic and intensity information. Across 10 datasets, their combination outperforms state-of-the-art LLIE methods.

  • HVI uses polarized HS maps and a learnable intensity component to address color bias and brightness artifacts.CIDNet models decoupled chromatic and intensity information in HVI for photometric adjustment under varying lighting conditions.
  • Experiments on 10 datasets show that HVI combined with CIDNet outperforms state-of-the-art LLIE methods.

Supplementary Material

Figure 7 visualizes the effect of the density-k parameter, with Intensity (Imax) as the independent variable and Ck as the dependent variable.

  • Figure 7 plots Ck against Intensity (Imax) for different density-k values.

7. Supplementary Introduction

The supplementary material extends the HVI-CIDNet analysis beyond the main text. It adds color-space details, generalization experiments, module analysis, and limitation discussion.

  • The appendix provides additional HVI color-space details and an extension addressing cross-dataset LLIE challenges.
  • It analyzes CIDNet's LCA module and its submodules through detailed structural descriptions and ablation studies.
  • Additional HVI-CIDNet experiments validate the advantages of the proposed color space and network.
  • The appendix also analyzes limitations of the method.

8. Details and Extensions of HVI Color Space

This section analyzes how density-k controls HVI's low-light geometry and extends HVI to cross-dataset generalization. The extension adapts hue and saturation mappings to camera and scene differences, while experiments report improved robustness.

  • 8.1. Visualization and Further Discussion of Parameter k: As density-k increases, noise amplification makes the conflict between image detail and noise increasingly apparent.
  • 8.1. Visualization and Further Discussion of Parameter k: Ck remaps Intensity from zero to one, with k controlling its gradient and therefore the radius of the HV plane across intensities.
  • 8.1. Visualization and Further Discussion of Parameter k: The trainable density-k parameter regulates low-light signal-to-noise ratio and can vary across networks and datasets.Larger k values intensify the conflict between noise and detail and increase color deviation, saturation, and contrast.
  • 8.1. Visualization and Further Discussion of Parameter k: The sinusoidal collapse formula is selected to reduce risks of gradient explosion and gradient vanishing during training.
  • 8.2. Further Develop of improving generalization between different datasets: Cross-dataset generalization is addressed with trainable hue transformation Pγ and saturation mapping T to accommodate camera and scene differences.Pγ maps Hue through a segmented linear function controlled by γG and γB, while T establishes saturation correspondences between scenes.
  • 8.2. Further Develop of improving generalization between different datasets: For an unknown camera, the appropriate γG and γB values cannot be determined, limiting the generalization of these parameters.
  • 8.2. Further Develop of improving generalization between different datasets: Adding T(·) and Pγ significantly improves PSNR, SSIM, and LPIPS in cross-dataset ablation experiments.

9. Architecture Details of CIDNet

CIDNet uses a dual-branch Lighten Cross-Attention module to separately process intensity and chromatic information, then exchange complementary features. Its IEL and CDL apply distinct intensity-enhancement and color-denoising operations, whose ablation improves performance when retained.

  • Lighten Cross-Attention: The LCA module combines a Cross Attention Block, Intensity Enhance Layer, and Color Denoise Layer across intensity and HV branches.The intensity branch uses IEL, while the HV branch uses CDL; CAB connects complementary branch information.
  • Lighten Cross-Attention: Cross attention uses one branch as query and the opposite branch as key and value to refine branch-specific features with complementary information.This design forces each branch to learn from the other branch during processing.
  • Intensity Enhance Layer: IEL decomposes intensity features into illumination and reflectance components, then reconstructs the enhanced output with residual connections.Its structure follows Retinex theory and uses element-wise multiplication with depth-wise convolutions.
  • Color Denoise Layer: CDL decomposes photometric features into wavelength and saturation, estimates their changes, and re-stacks the components to denoise color features.The process uses depth-wise 1 × 1 and grouped 3 × 3 convolutions, followed by point-wise reconstruction.
  • Ablation: Removing CAB, IEL, or CDL decreases PSNR and SSIM on LOLv2-Real, while removing CAB causes unstable brightness and removing IEL or CDL produces overly dark results.The full LCA design therefore supports both brightness stability and detail preservation.

10. Additional Experiments and Details

Additional experiments evaluate CIDNet across paired, unpaired, large-scale, and updated LOL benchmarks. The results emphasize competitive quality and efficiency, while also identifying a parameter-fitting limitation in some comparison methods.

  • Experiment Scope: CIDNet is evaluated on LOL, five unpaired datasets, SICE, SID, LOL-Blur, and SIDD to test enhancement, generalization, deblurring, and denoising.The dataset suite spans standard, unpaired, large-scale, joint, and real-noise settings.
  • Evaluation Details: GT Mean adjusts output brightness toward GroundTruth during evaluation to reduce luminance fluctuations on the small LOLv1 test set.The adjustment is intended to compare non-luminance attributes more precisely.
  • Evaluation Details: CIDNet combines HVI-space and sRGB-space losses using L1, SSIM, edge, and perceptual components with balancing weights.The loss formulation integrates objectives from both color representations.
  • Updated Comparisons: CIDNet achieves competitive updated-method metrics with the lowest FLOPs, although GLARE has a slightly better LPIPS score on LOLv2-Real.The paper attributes GLARE’s LPIPS advantage to smoother outputs that sacrifice finer details.
  • Updated Comparisons: Compared with GLARE, CoLIE, Zero-IG, and LightenDiff, the paper links weaker fitting of precise results to substantially smaller parameter counts in CoLIE and Zero-IG.The authors identify this as a reason for pursuing methods combining visual quality and generalization.

10.4. Joint Low-light Image Deblurring and Enhancement

CIDNet is tested on joint low-light enhancement and deblurring, real denoising, and additional HVI analyses. It improves quantitative and visual recovery in difficult settings, while experiments also examine HVI components and generalization preprocessing.

  • Joint Deblurring and Enhancement: CIDNet produces more recognizable LOL-Blur reconstructions with realistic colors, sharper edges, and improved visual comfort compared with competing methods.The qualitative comparisons include GroundTruth references.
  • Real Noise Removal: CIDNet reaches 39.88 dB PSNR on SIDD, 0.12 dB above MIRNet, supporting its evaluation on real image denoising.The denoising results are reported in Table 11 and Figure 18.
  • HVI Analysis: The HVI framework separates intensity and HV information, while polarization reduces red-region noise and the intensity-collapse function addresses dark-plane density.Error-map comparisons are used to examine these color-space effects.
  • Generalization: Random gamma preprocessing with γ between 0.6 and 1.2 reduces NIQE by 0.389 and is used to improve generalization on five unpaired datasets.Because competing methods did not use this technique, the paper avoids direct comparisons for this result.

11. Limitation and Unstudied issues

The paper identifies unresolved questions concerning the intensity-collapse function, transfer beyond LLIE, training paradigms, and separate HVI/CIDNet optimization. It also provides qualitative evidence across several datasets, including extremely dark and blurred scenes.

  • Intensity Collapse Function: The authors note that whether a more appropriate function than the sine-based intensity-collapse function exists remains unexplored.The current function collapses dark regions in the HVI color-space domain.
  • Other Tasks: Applying HVI to SwinIR super-resolution improves average PSNR by 0.14 dB, but broader applicability to other vision tasks remains unverified.The authors cite time constraints and propose future evaluation on additional tasks.
  • Training Paradigms: CIDNet is validated only with supervised learning; unsupervised, semi-supervised, and zero-shot training remain unexplored.These paradigms are identified as future research directions.
  • Separate Training: HVI and CIDNet cannot currently be trained separately because ground-truth HVI representations are unavailable.The paper leaves the performance impact of separate training unresolved.
  • Qualitative Evaluation: On extremely low-light Sony-Total-Dark images, competing methods either extract few features or introduce artifacts and color bias, whereas CIDNet produces a more realistic enhancement.Additional qualitative comparisons cover LOL subsets, unpaired datasets, and LOL-Blur.
Loading 2502.20272v2…