RHVI-FDD: A Hierarchical Decoupling Framework for Low-Light Image Enhancement

Junhao Yang; Bo Yang; Hongwei Ge; Yanchun Liang; Heow Pueh Lee; Chunguo Wu

cs.CVApr 7, 2026

RHVI-FDD: A Hierarchical Decoupling Framework for Low-Light Image Enhancement

Authors: Junhao Yang, Bo Yang, Hongwei Ge, Yanchun Liang, Heow Pueh Lee, Chunguo Wu

Organizations: Jilin University, Changchun, China · Dalian University of Technology, Dalian, China · National University of Singapore, Singapore

Abstract

Low-light images often suffer from severe noise, detail loss, and color distortion, which hinder downstream multimedia analysis and retrieval tasks. The degradation in low-light images is complex: luminance and chrominance are coupled, while within the chrominance, noise and details are deeply entangled, preventing existing methods from simultaneously correcting color distortion, suppressing noise, and preserving fine details. To tackle the above challenges, we propose a novel hierarchical decoupling framework (RHVI-FDD). At the macro level, we introduce the RHVI transform, which mitigates the estimation bias caused by input noise and enables robust luminance-chrominance decoupling. At the micro level, we design a Frequency-Domain Decoupling (FDD) module with three branches for further feature separation. Using the Discrete Cosine Transform, we decompose chrominance features into low, mid, and high-frequency bands that predominantly represent global tone, local details, and noise components, which are then processed by tailored expert networks in a divide-and-conquer manner and fused via an adaptive gating module for content-aware fusion. Extensive experiments on multiple low-light datasets demonstrate that our method consistently outperforms existing state-of-the-art approaches in both objective metrics and subjective visual quality.

Figures & tables

Figure 1. Methods overview and visual comparison. (a) Conventional methods, which often produce color bias and brightness artifacts due to inherent high color sensitivity in sRGB. (b) HVI mitigates color shifts via luminance–chrominance decoupling, but coarse illumination estimation under noise induces overexposure, uneven brightness, and entangled chrominance features. (c) Ours: employs the optimized RHVI transformation for more robust luminance-chrominance decoupling, and applies feature-level decoupling in the chrominance map to separately handle tone, details, and noise. Comparison of three methods. (a) shows sRGB methods with color artifacts. (b) shows HVI methods with noise issues. (c) shows the proposed RHVI method with clean details and balanced colors.

Figure 2. (I) RHVI refines noise-corrupted chrominance in HVI via IRM, achieving robust luminance-chrominance decoupling and generating luminance and chrominance maps, which are processed by a dual-branch network for the feature refinement. (II) The FDD framework, embedded in the chrominance branch, projects features into the frequency domain via DCT and splits them into three bands for hue-structure, detail, and noise. Each band is handled by expert networks (GCM, DRG, ANSU) and fused by the fusion module (ACGF). Diagram of the proposed method. Part I shows the RHVI transform process decoupling luminance and chrominance. Part II shows the FDD framework processing features in the frequency domain with three distinct branches for hue, detail, and noise, followed by a fusion module.

Figure 3. Comparison of luminance maps. (a) Normal-light image. (b) Luminance map produced by the luminance branch of HVI. (c) Luminance map obtained by feeding the RHVI-transformed result into the same branch.

Figure 4. Comparison of chrominance maps. (a) Low-light image. (b) Chrominance map of (a) after HVI decoupling. (c) Chrominance map of the ground-truth image processed similarly. Both (b) and (c) are visualized using the same pseudocolor reconstruction.

Figure 5. Detailed architectures of the proposed modules: (a) Illumination Refinement Module (IRM), (b) Global Context Modulator (GCM), (c) Detail Refinement Gated Block (DRG), (d) Adaptive Noise Suppression Unit (ANSU), and (e) Adaptive Channel-Guided Fusion (ACGF). Schematic diagrams of the five proposed modules. (a) shows the Illumination Refinement Module. (b) shows the Global Context Modulator. (c) shows the Detail Refinement Gated Block. (d) shows the Adaptive Noise Suppression Unit. (e) shows the Adaptive Channel-Guided Fusion module.

Panel of Figure 6 — Figure 6. Spatial reconstructions of downsampled bottleneck features, obtained by projecting individual frequency bands back to the spatial domain via IDCT, revealing distinct frequency-band responses and supporting the underlying frequency decoupling assumption.

Methods	Reference	LOLv1			LOLv2-Real			LOLv2-Synthetic			Complexity
Methods	Reference	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Params/M	FLOPs/G
RetinexNet ( Wei et al., 2018 )	BMVC’18	18.915	0.427	0.470	16.097	0.401	0.543	17.137	0.762	0.255	0.84	584.47
KinD ( Zhang et al., 2019 )	MM’19	23.018	0.843	0.156	17.544	0.669	0.375	18.320	0.796	0.252	8.02	34.99
MIRNet ( Zamir et al., 2020 )	ECCV’20	24.140	0.842	0.131	20.357	0.782	0.317	21.927	0.833	0.209	31.76	785
RUAS ( Liu et al., 2021a )	CVPR’21	18.654	0.518	0.270	15.326	0.488	0.310	13.765	0.638	0.305	0.003	0.83
LLFlow ( Wang et al., 2022 )	AAAI’22	24.998	0.871	0.117	17.433	0.831	0.176	24.807	0.919	0.067	17.42	358.4

Table 1. Quantitative results of PSNR

\uparrow

, SSIM

\uparrow

, and LPIPS

\downarrow

on the LOL (v1 and v2) datasets. The FLOPs is tested on a single

256\times 256

image. The best results are marked in red and the second-best results are marked in cyan respectively. Please note that we apply the GT mean method only on the LOLv1 to minimize errors due to its limited size.

Panel of Figure 7 — Figure 7. Visual comparison of the enhanced images yielded by different SOTA methods on LOLv1 (upper) and LOLv2 (lower). A multi-panel figure comparing image enhancement results. The top row shows results on LOLv1 dataset, and the bottom row shows results on LOLv2. Columns represent Input, KinD, MIRNet, SNR, CIDNet, UHDFormer, Ours, and Ground Truth. Our method shows better detail retention and color balance.

Methods	SICE-Grad		SICE-Mix		Sony-Total-Dark
Methods	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$
RetinexNet ( Wei et al., 2018 )	12.450	0.619	12.397	0.606	15.695	0.395
Zero-DCE ( Guo et al., 2020 )	12.475	0.644	12.428	0.633	14.087	0.090
RUAS ( Liu et al., 2021a )	8.628	0.494	8.684	0.493	12.622	0.081
URetinexNet ( Wu et al., 2022 )	10.894	0.610	10.903	0.600	12.622	0.081
LLFlow ( Wang et al., 2022 )	12.737	0.617	12.737	0.617	16.226	0.367

Table 2. Quantitative comparison on SICE (SICE-Grad, SICE-Mix) and Sony-Total-Dark datasets.

Methods	DICM	LIME	MEF	NPE	VV
Zero-DCE ( Guo et al., 2020 )	4.580	5.820	4.930	4.530	4.810
KinD ( Zhang et al., 2019 )	3.614	4.772	4.819	4.175	3.835
FECNet ( Huang et al., 2022 )	4.139	6.041	4.707	4.500	3.346
MIRNet ( Zamir et al., 2020 )	4.042	6.453	5.504	5.235	4.735
SMG ( Xu et al., 2023 )	4.733	5.451	5.754	5.208	4.884
Bread ( Guo and Hu, 2023 )	4.179	4.717	5.369	4.160	3.304

Table 3. Quantitative comparison (NIQE

\downarrow

) on five real-world unpaired datasets (DICM, LIME, MEF, NPE, and VV).

Methods	RetinexNet ( Wei et al., 2018 )	LEDNet ( Wang et al., 2019 )	SNR-Aware ( Xu et al., 2022b )	LLFormer ( Wang et al., 2023b )	LCDBNet ( Wang et al., 2024b )	GSAD ( Hou et al., 2023 )	CIDNet ( Yan et al., 2025 )
PSNR $\uparrow$	20.803 ( +4.789 )	23.307 ( +3.529 )	22.017 ( +0.697 )	22.710 ( +2.678 )	24.029 ( +3.682 )	23.821 ( +3.668 )	24.822 ( +0.711 )
SSIM $\uparrow$	0.652 ( +0.108 )	0.838 ( +0.031 )	0.839 ( -0.006 )	0.866 ( +0.074 )	0.833 ( +0.053 )	0.874 ( +0.029 )	0.882 ( +0.011 )
LPIPS $\downarrow$	0.351 ( -0.106 )	0.113 ( -0.013 )	0.111 ( -0.067 )	0.121 ( -0.090 )	0.126 ( -0.048 )	0.101 ( -0.012 )	0.078 ( -0.030 )
Params / M	0.84	7.4	4.01	24.55	7.36	17.36	1.88
Model Type	CNN	CNN	Transformer	Transformer	CNN	Diffusion	Transformer
Branch	Dual	Single	Single	Single	Dual	Single	Dual

Table 4. Results of applying RHVI-FDD as a plug-in to various LLIE methods on the LOLv2-Real dataset. Values in brackets represent the absolute performance gain after integration. Red values indicate performance improvement (corresponding to higher PSNR/SSIM

\uparrow

or lower LPIPS

\downarrow

), while green values indicate a performance drop. Our RHVI-FDD has 0.63M parameters.

Methods	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Ours (RHVI $\to$ HSV)	21.698	0.822	0.169
Ours (RHVI $\to$ HVI)	23.921	0.867	0.120
Ours w/o IRM	24.326	0.871	0.115
Ours w/o GCM (Low-freq)	23.952	0.865	0.110
Ours w/o DRG (Mid-freq)	23.518	0.852	0.132
Ours w/o ANSU (High-freq)	24.215	0.873	0.098

Table 5. Ablation study on different components of our framework on LOLv2-Real.

Panel of Figure 8 — Figure 8. Ablation study on the contributions of low, mid, and high-frequency components. A grid comparison showing the ablation study results of our hierarchical decoupling paradigm. (a) is the underexposed and noisy input. (b) without the low-frequency component shows insufficient brightness and poor global tone. (c) without the mid-frequency component exhibits blurred structures and a loss of fine textures. (d) without the high-frequency component shows significant residual noise and artifacts. (e) is the full model, which achieves natural illumination and sharp details comparable to the ground truth (f).

Explore similar work

Frequency-Decomposed INR for NIR-Assisted Low-Light RGB Image Denoising
Apr 18, 2026Ligen Shi, Zengyu Pang, Chang Liu +2Implicit Neural Representations Frequency-Domain Image Processing
College of Computer Science (College of Software), Inner Mongolia University, Hohhot, 010021, China · School of Mathematics and Statistics, Wuhan University, Wuhan, 430072, China · Institute of Computational Imaging, Beijing Information Science and Technology University, Beijing 102206, China
Low-light Image Enhancement via Multi-scale Attention combined with Fourier Transform
Jul 27, 2026Wenbin Du, Jian Long, Zhu CaoFrequency-Domain Feature Learning Low-Light Image Enhancement
aKey Laboratory of Smart Manufacturing in Energy Chemical Process, Ministry of Education, East China University of Science and Technology, Shanghai, 200237, China · bCollege of Electronics and Information Engineering, Tongji University, Shanghai 201804, China · cShanghai Research Institute for Intelligent Autonomous Systems, Tongji University, Shanghai 201210, China
IDM-Net: A Lightweight Illumination-Decoupled Modulation Network for Low-Light Image Enhancement
Sep 25, 2026Cheng-Yen Hsiao, Jing-Ming GuoLow-Light Image Enhancement
Department of Electrical Engineering National Taiwan University of Science and Technology Taipei, Taiwan

Apr 18, 2026J/K move · Enter open · S save

Frequency-Decomposed INR for NIR-Assisted Low-Light RGB Image Denoising

Ligen Shi, Zengyu Pang, Chang Liu, Shuchen Sun+1

College of Computer Science (College of Software), Inner Mongolia University, Hohhot, 010021, China · School of Mathematics and Statistics, Wuhan University, Wuhan, 430072, China · Institute of Computational Imaging, Beijing Information Science and Technology University, Beijing 102206, China

Implicit Neural Representations Frequency-Domain Image Processing Image Restoration Image Denoising

Open paper AskPDF arXiv

Addressing the issues of severe noise and high frequency structural degradation in visible images under low-light conditions, this paper proposes a Near Infrared (NIR) aided low light image restoration method based on Frequency Decoupled Implicit Neural Representation (FDINR). Based on the statistical prior of RGB-NIR cross-modal frequency correlations, specifically that low-frequency RGB signals are more reliable, whereas high frequency NIR signals exhibit higher correlation, we explicitly decompose images into distinct frequency components via multi-scale wavelet transforms and construct a dual-branch implicit neural representation framework. Within this framework, we design a cross modal differentiated frequency supervision mechanism, leveraging low light RGB to guide the reconstruction of low frequency luminance and color, and utilizing high-SNR NIR signals to constrain the generation of high frequency texture details, thereby achieving complementary advantages in the frequency domain. Furthermore, an uncertainty-based adaptive weighting loss function is introduced to automatically balance the contributions of different frequency tasks, solving the problems of color distortion and artifacts caused by rigid fusion in the spatial domain common in traditional methods. Experimental results demonstrate that FD-INR not only effectively restores image luminance consistency and structural details but also, benefitting from its implicit continuous representation, outperforms existing methods in arbitrary-resolution reconstruction tasks, significantly enhancing the reliability of low light perception.

Abstract

Explore similar work

Frequency-Decomposed INR for NIR-Assisted Low-Light RGB Image Denoising

Low-light Image Enhancement via Multi-scale Attention combined with Fourier Transform

IDM-Net: A Lightweight Illumination-Decoupled Modulation Network for Low-Light Image Enhancement