PRIN: A Perceptual-Structural Retinex-Inspired Network for Low-Light Image Enhancement

Qiyi Xie; Wei Chen Ng; Ziyao Li; Yuning Wu; Huasen Wu; Tianshi Miao; Jian Ding Tan

eess.IVOct 30, 2026

PRIN: A Perceptual-Structural Retinex-Inspired Network for Low-Light Image Enhancement

Authors: Qiyi Xie, Wei Chen Ng, Ziyao Li, Yuning Wu, Huasen Wu, Tianshi Miao, Jian Ding Tan

Organizations: School of Artificial Intelligence and Robotics, Xiamen University Malaysia, Jalan Sunsuria, Bandar Sunsuria, Sepang, 43900, Selangor, Malaysia · School of Engineering and Applied Science, University of Pennsylvania, 220 South 33rd Street, Philadelphia, 19104, PA, United States · S.M.A.R.T. Nexus Center of Excellence, Xiamen University Malaysia, Jalan Sunsuria, Bandar Sunsuria, Sepang, 43900, Selangor, Malaysia

Abstract

Low-light image enhancement (LLIE) focuses not only on visibility recovery but also on preserving structurally meaningful content and perceptual naturalness. Existing learning-based methods may achieve favorable pixel-wise fidelity while still producing over-smoothed textures, color deviations, or spatially incoherent illumination. Retinex-inspired methods provide interpretable decomposition into reflectance and illumination components, however, the asymmetric restoration roles of these components are rarely encoded through branch-specific frequency-domain refinement. To address this issue, we propose a Perceptual-Structural Retinex-Inspired Network (PRIN), which integrates Retinex priors with restoration-role-aware enhancement. PRIN employs a shared encoder and asymmetric decoders to estimate reflectance and illumination components. The reflectance branch performs attention-guided and edge-aware wavelet refinement to preserve textures and boundaries, whereas the illumination branch promotes spatially coherent brightness adjustment. After the recomposition of reflectance and illumination branches, a curve-based post-composition enhancement head calibrates residual tone and exposure deviations. Experiments on LOL-v1, LOL-v2-Syn, and LOL-v2-Real show that the proposed network ties for the best structural similarity index (SSIM) and obtains the second-lowest learned perceptual image patch similarity (LPIPS) on LOL-v1, achieves the best SSIM and LPIPS on LOL-v2-Syn, and obtains the highest peak signal-to-noise ratio (PSNR) and the second-lowest LPIPS on LOL-v2-Real.

Figures & tables

Panel of Figure 1 — Figure 1: The decomposition result of Haar wavelet transform.

Panel of Figure 2 — Figure 2: Visual comparison between the original enhanced result and its blur-based post-processing variants. The blur variants are generated by applying Gaussian smoothing with different radii $r$ to the original output and then blending the smoothed image with the original result using a fixed weight $\alpha=0.2$ .

Method	Modeling Principle	Restoration Focus	Frequency / Calibration Design
Retinex-Net [ 36 ]	Learned Retinex decomposition	Reflectance-illumination decomposition and illumination enhancement	No explicit frequency refinement
KinD [ 46 ]	Retinex restoration with degradation modeling	Reflectance restoration and adjustable illumination enhancement	No explicit frequency refinement
SGM-Net [ 42 ]	Sparse-gradient Retinex restoration	Gradient-regularized illumination estimation and component enhancement	Structure regularization without wavelet refinement
URetinex-Net [ 38 ]	Unfolded Retinex optimization	Iterative decomposition and restoration learning	No explicit frequency refinement
Retinexformer [ 3 ]	Illumination-guided Retinex Transformer	Illumination-guided global feature restoration	Attention-based long-range modeling
Diff-Retinex [ 43 ]	Diffusion-based Retinex restoration	Component-wise generative restoration	Diffusion refinement for reflectance and illumination

Table 1: Technical comparison of representative low-light image enhancement methods and the proposed PRIN.

Mechanism	Role prior it encodes	Realisation
Asymmetric branch design	Illumination is globally consistent and low-frequency; reflectance is not.	The illumination decoder applies three additional down-sampling stages, reaching a spatial resolution of $1/32$ of the input and thereby providing a larger effective receptive field; the reflectance decoder does not.
Component-specific wavelet refinement	The two factors require different sub-band treatment.	WDGRR refines reflectance sub-bands with a detail-oriented path; WSAIR refines illumination sub-bands with a structure-oriented path.
Bounded output projections	The scaling freedom of Eq. (2) must be removed.	Sigmoid-bounded branch outputs constrain $\hat{R}$ and $\hat{L}$ to a feasible range.
Loss-based regularization	Illumination should be piecewise smooth; the product must stay faithful; sub-bands are constrained band-wise.	$\mathcal{L}_{\text{illum}}$ , an edge-aware weighted total variation (TV) term applied to $\hat{L}$ only (§3.7.6); $\mathcal{L}_{\text{base}}$ , product consistency on $\hat{I}_{\text{pre}}$ (§3.7.5); and $\mathcal{L}_{\text{wav}}$ , sub-band weighting.

Table 2: Mechanisms used to encode restoration-role awareness in the proposed decomposition framework.

Figure 3: The overall architecture of the proposed PRIN

Method	Reference	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$
MIRNet [ 44 ]	ECCV’20	0.820	0.160	23.480
EnlightenGAN [ 14 ]	TIP’21	0.651	0.322	17.600
SGM-Net [ 42 ]	TIP’21	0.816	0.169	19.370
CDEF [ 17 ]	TMM’22	0.585	0.407	16.335
SCI [ 25 ]	CVPR’22	0.526	0.392	14.784
SNR-Net [ 39 ]	CVPR’22	0.842	0.233	24.613

Table 3: Quantitative comparison on the LOL dataset. The best and second-best results are highlighted in bold and underlined , respectively.

Methods	Reference	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$
Zero-DCE [ 7 ]	CVPR’20	0.815	0.140	17.710
MIRNet [ 44 ]	ECCV’20	0.876	0.122	21.940
RUAS [ 22 ]	CVPR’21	0.652	0.335	16.550
SCI [ 25 ]	CVPR’22	0.801	0.301	22.203
SNR-Net [ 39 ]	CVPR’22	0.912	0.187	24.142
Uformer [ 34 ]	CVPR’22	0.871	0.060	22.037

Table 4: Quantitative comparison on the LOL-v2-Syn dataset. The best and second-best results are highlighted in bold and underlined , respectively.

Methods	Reference	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$
EnlightenGAN [ 14 ]	TIP’21	0.678	0.364	18.676
RUAS [ 22 ]	CVPR’21	0.626	0.395	15.351
SNR-Net [ 39 ]	CVPR’22	0.849	0.237	21.482
URetinex-Net [ 38 ]	CVPR’22	0.853	0.208	21.093
CDEF [ 17 ]	TMM’22	0.630	0.349	19.757
SCI [ 25 ]	CVPR’22	0.540	0.345	17.304

Table 5: Quantitative comparisons on LOL-v2-Real. The best and second-best results are highlighted in bold and underlined , respectively.

Panel of Figure 4 — Figure 4: Visual comparison of different low-light image enhancement methods on representative examples from the LOL-v1 dataset.

Panel of Figure 5 — Figure 5: Visual comparison of different low-light image enhancement methods on representative examples from the LOL-v2-Syn dataset.

Panel of Figure 6 — Figure 6: Visual comparison of different low-light image enhancement methods on representative examples from the LOL-v2-Real dataset.

Variant	Block Design	Down/Up sampling	Illumination	Wavelet Position	Post-Enhancement
V1a	Conv-BN-ReLU	MaxPool / ConvTranspose	1-channel	64-channel	GammaHead
V1b	BatchNorm+PReLU (PreAct)	MaxPool / ConvTranspose	3-channel	64-channel	GammaHead
V2	GroupNorm + SiLU	PixelUnshuffle / PixelShuffle	3-channel	64-channel	DCECurveEnhancer
V3	GroupNorm + SiLU	PixelUnshuffle / PixelShuffle	3-channel	128-channel	GlobalAffineGammaHead

Table 6: Comparison of different network variants.

Variant	Params (M)	LOL-v1			LOL-v2-Real			LOL-v2-Syn
Variant	Params (M)	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$
V1a	5.210	0.219	0.803	20.55	-	-	-	-	-	-
V1b	23.394	0.155	0.830	22.02	0.269	0.809	18.870	-	-	-
V2 (current)	24.412	0.119	0.853	23.390	0.181	0.852	22.428	0.056	0.931	24.596
V3	26.247	0.110	0.852	23.306	0.427	0.760	20.176	0.058	0.929	24.661

Table 7: Quantitative comparison of different network variants on three datasets. The best and second-best results are highlighted in bold and underlined , respectively.

Panel of Figure 7 — Figure 7: Qualitative comparison of different variants on representative samples from LOL-v1 and LOL-v2-Syn.

Method	Reference	LOL-v1			LOL-v2-Real			LOL-v2-Syn
Method	Reference	MUSIQ $\uparrow$	LIQE $\uparrow$	Q-Align $\uparrow$	MUSIQ $\uparrow$	LIQE $\uparrow$	Q-Align $\uparrow$	MUSIQ $\uparrow$	LIQE $\uparrow$	Q-Align $\uparrow$
KinD [ 46 ]	ACM MM’19	69.189	2.853	3.632	66.225	2.785	3.427	61.747	2.708	2.503
MIRNet [ 44 ]	ECCV’20	67.276	3.052	3.581	63.348	3.103	3.457	58.519	2.898	2.551
RUAS [ 22 ]	CVPR’21	59.345	2.902	3.341	56.177	3.077	3.269	51.735	2.006	1.876
Restormer [ 45 ]	CVPR’22	64.236	3.266	3.390	51.711	2.897	2.782	63.545	3.817	2.857
SNR-Net [ 39 ]	CVPR’22	65.223	2.552	3.424	62.506	2.965	3.393	66.430	3.883	3.324

Table 8: Non-reference image quality assessment results on LOL-v1, LOL-v2-Real, and LOL-v2-Syn. The best and second-best results are highlighted in bold and underlined , respectively.

Method	Weight	DICM			LIME			MEF
Method	Weight	MUSIQ $\uparrow$	LIQE $\uparrow$	Q-Align $\uparrow$	MUSIQ $\uparrow$	LIQE $\uparrow$	Q-Align $\uparrow$	MUSIQ $\uparrow$	LIQE $\uparrow$	Q-Align $\uparrow$
Input	–	58.019	3.153	3.734	58.277	2.808	3.341	57.536	2.889	3.083
Zero-DCE [ 7 ]	Official	58.857	3.182	3.517	60.734	3.063	3.107	64.925	3.786	3.550
RUAS [ 22 ]	Official	53.502	2.058	2.536	58.019	2.628	2.670	61.805	3.180	3.209
SCI [ 25 ]	Official	53.596	2.536	2.909	59.113	2.891	2.961	62.532	3.527	3.435
URetinex-Net [ 38 ]	LOL	61.146	3.329	3.679	64.297	3.269	3.369	67.273	3.923	3.691

Table 9: Zero-shot cross-dataset evaluation on DICM, LIME, and MEF. Higher values indicate better performance for all reported metrics. The best result in each column is highlighted in bold.

Method	Weight	NPE			VV
Method	Weight	MUSIQ $\uparrow$	LIQE $\uparrow$	Q-Align $\uparrow$	MUSIQ $\uparrow$	LIQE $\uparrow$	Q-Align $\uparrow$
Input	–	65.131	3.853	3.735	43.600	1.468	3.485
Zero-DCE [ 7 ]	Official	68.964	3.890	3.298	39.875	1.296	3.564
RUAS [ 22 ]	Official	61.430	2.411	2.145	35.172	1.051	2.692
SCI [ 25 ]	Official	62.290	3.101	2.556	32.998	1.141	3.006
URetinex-Net [ 38 ]	LOL	68.114	3.839	3.527	40.257	1.521	3.073

Table 10: Zero-shot cross-dataset evaluation on NPE and VV. Higher values indicate better performance for all reported metrics. The best result in each column is highlighted in bold.

Panel of Figure 8 — Figure 8: Qualitative comparison on representative images from the unpaired real-world benchmarks. The three rows show scenes with challenging illumination transitions and large differences between shadow and highlight regions.

Baseline	WDGRR	WSAIR	DCE	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$
✓				0.269	0.809	18.870
✓	✓			0.178	0.845	21.398
✓		✓		0.211	0.824	20.257
✓			✓	0.201	0.832	20.841
✓	✓	✓		0.197	0.834	20.840
✓	✓	✓	✓	0.181	0.852	22.428

Table 11: Ablation study of the proposed modules on LOL-v2-Real.

\checkmark

indicates that the corresponding component is enabled.

Variant	PSNR $\uparrow$	SSIM $\uparrow$	MAE $\downarrow$	LPIPS $\downarrow$
Full objective	23.36	0.8439	0.0712	0.1293
w/o $\mathcal{G}_{1}$ (structure-edge)	22.68	0.8065	0.0673	0.1865
w/o $\mathcal{G}_{2}$ (frequency-domain)	23.37	0.8507	0.0684	0.1326
w/o $\mathcal{G}_{3}$ (decomposition regularization)	23.23	0.8446	0.0707	0.1315
w/o $\mathcal{G}_{4}$ (color-exposure)	23.26	0.8519	0.0688	0.1259

Table 12: Loss-group ablation results on the LOL-v1 test set.

\mathcal{L}_{\mathrm{rec}}

is retained in all configurations.

Measurement	$\hat{R}$	$\hat{L}$	$p$
Normalized gradient energy	0.0254	0.0096	$6\times 10^{-5}$
Correlation with ground-truth edges	0.875	0.431	$6\times 10^{-5}$
Illumination invariance $\rho(X_{\mathrm{low}},X_{\mathrm{ref}})$	0.977	0.743	$6\times 10^{-5}$

Table 13: Quantitative evidence of behavioral separation between the reflectance and illumination branches.

Config	$(s_{R},s_{L},s_{E})$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
C0 default	$(0.10,0.05,0.10)$	22.620	0.8441	0.1388
C1 ratio 1:1	$(0.10,0.10,0.10)$	23.377 (0.107)	0.8520 (0.022 ∗ )	0.1252 (0.0006 ∗∗ )
C2 ratio 1:2	$(0.05,0.10,0.10)$	23.056 (0.083)	0.8497 (0.010 ∗ )	0.1243 (0.0003 ∗∗ )
C3 magnitude $0.5\times$	$(0.05,0.025,0.10)$	23.017 (0.454)	0.8456 (0.252)	0.1294 (0.0026 ∗∗ )
C4 magnitude $2\times$	$(0.20,0.10,0.10)$	22.814 (0.454)	0.8451 (0.303)	0.1269 (0.0004 ∗∗ )
C5 edge $0.5\times$	$(0.10,0.05,0.05)$	22.729 (0.561)	0.8432 (0.600)	0.1276 (0.0067 ∗∗ )

Table 14: Sensitivity of PRIN to the residual-scale coefficients on LOL-v1.

s_{R}

s_{L}

, and

s_{E}

denote the reflectance-refinement, illumination-refinement, and edge-injection scales, respectively. Values in parentheses report two-sided paired Wilcoxon

p

-values against C0 over the 15 test images.

{}^{*}p<0.05

and

{}^{**}p<0.01

Quantity	Agreement	Relative disagreement
Enhanced output $\hat{I}$	0.9825	—
Reflectance $\hat{R}$	0.9710	$1.7\times$
Illumination $\hat{L}$	0.7797	$\mathbf{12.6\times}$

Table 15: Cross-configuration agreement of the enhanced output and the recovered Retinex factors.

Dataset	PSNR $\uparrow$			SSIM $\uparrow$			LPIPS $\downarrow$
Dataset	w/o	w/	$\Delta$	w/o	w/	$\Delta$	w/o	w/	$\Delta$
LOL-v1	23.390	27.055	+3.666	0.853	0.872	+0.018	0.119	0.114	-0.005
LOL-v2-Real	22.428	27.019	+4.591	0.852	0.876	+0.023	0.181	0.170	-0.011
LOL-v2-Syn	24.596	28.610	+4.014	0.931	0.942	+0.011	0.056	0.054	-0.002

Table 16: Effect of GT-mean correction on quantitative results.

Method	Params (M) $\downarrow$	FLOPs@ $256\times 256$ (G) $\downarrow$	FLOPs@ $600\times 400$ (G) $\downarrow$	Latency (ms) $\downarrow$	FPS $\uparrow$
MIRNet [ 44 ]	31.79	511.75	1874.08	294.2	3.4
SNR-Net [ 39 ]	39.12	23.96	88.91	18.8	53.3
Uformer-T [ 34 ]	5.20	9.98	62.37	79.3	12.6
Restormer [ 45 ]	26.10	140.99	516.32	269.0	3.7
LLFormer [ 30 ]	24.52	22.03	81.77	371.9	2.7
Retinexformer [ 3 ]	1.61	15.76	57.72	36.0	27.8

Table 17: Comparison of computational complexity and practical inference efficiency.

Figure 10: SSIM-LPIPS performance distribution on LOL-v1.

Image Enhancement

Image enhancement restores visibility, color, detail, and dynamic range in degraded images for human viewing and downstream vision tasks.

See all 156 papers

WREN: Low Light Image Enhancement Using Retinex theory-based Double U-Net-like Structures
Jul 17, 2026Reina Kaneko, Junya Hara, Hiroshi Higashi +1Image Enhancement U-Net
The University of Osaka, Osaka, Japan · Kansai University, Osaka, Japan
InterLight: Leveraging Intrinsic Illumination Priors for Low-Light Image Enhancement
May 19, 2026Ziqi Wang, Xu Zhang, Laibin Chang +3Image Enhancement Low-Light Image Enhancement
National Engineering Research Center for Multimedia Software, School of Computer Science, Wuhan University · Department of Computer Science, University of Macau · Department of Computer Vision, Mohamed bin Zayed University of Artificial Intelligence +1
Multinex: Lightweight Low-light Image Enhancement via Multi-prior Retinex
Apr 11, 2026Alexandru Brateanu, Tingting Mu, Codruta Ancuti +1Low-Light Image Enhancement Image Enhancement
Department of Computer Science, University of Manchester, Manchester, United Kingdom · ETcTI, University Politehnica Timisoara, Timisoara, Romania · West University of Timisoara, Timisoara, Romania

Jul 17, 2026J/K move · Enter open · S save

WREN: Low Light Image Enhancement Using Retinex theory-based Double U-Net-like Structures

Reina Kaneko, Junya Hara, Hiroshi Higashi, Yuichi Tanaka

The University of Osaka, Osaka, Japan · Kansai University, Osaka, Japan

Image Enhancement U-Net Low-Light Image Enhancement

Open paper AskPDF arXiv

This paper proposes a neural network for low light image enhancement (LLIE) based on retinex theory to make LLIE robust for various dynamic range scenes. The retinex theory is an image formulation model inspired by a human color perception hypothesis, where a low light image is decomposed into intrinsic color context (i.e., reflectance map) and scene-dependent illumination (i.e., illumination map). Due to non-uniqueness of its decomposition, existing retinex-based LLIE methods often fail to achieve stable decomposition, which lead to over-enhancement. Typically, they are sensitive to the dynamic ranges that vary in different lighting conditions. To tackle this issue, we propose WREN: An LLIE neural network with double U-Net-like structures. WREN consists of two U-Net-like sub-networks. The first network has one encoder and two decoders that decompose an input image into the reflectance and illumination maps. The second network with a customized Transformer block between an encoder and a decoder only enhances the illumination map obtained from the first network: This completely follows the assumption of the retinex theory. Finally, the enhanced illumination map is recombined with the reflectance map. The network is trained end-to-end with a scale-invariant loss function, which gives robustness against the illumination scaling. Numerical results show that our method achieves the state-of-the-art performance across multiple datasets. Our code is available online.

Abstract

Image Enhancement

Explore similar work

WREN: Low Light Image Enhancement Using Retinex theory-based Double U-Net-like Structures

InterLight: Leveraging Intrinsic Illumination Priors for Low-Light Image Enhancement

Multinex: Lightweight Low-light Image Enhancement via Multi-prior Retinex