Rethinking Exposure Correction for Spatially Non-uniform Degradation

Ao Li; Jiawei Sun; Le Dong; Zhenyu Wang; Weisheng Dong

cs.CVApr 5, 2026

Rethinking Exposure Correction for Spatially Non-uniform Degradation

Authors: Ao Li, Jiawei Sun, Le Dong, Zhenyu Wang, Weisheng Dong

Organizations: School of Artificial Intelligence, Xidian University, Xi’an, China · Hangzhou Institute of Technology, Xidian University, Hangzhou, China

Abstract

Real-world exposure correction is fundamentally challenged by spatially non-uniform degradations, where diverse exposure errors frequently coexist within a single image. However, existing exposure correction methods are still largely developed under a predominantly uniform assumption. Architecturally, they typically rely on globally aggregated modulation signals that capture only the overall exposure trend. From the optimization perspective, conventional reconstruction losses are usually derived under a shared global scale, thus overlooking the spatially varying correction demands across regions. To address these limitations, we propose a new exposure correction paradigm explicitly designed for spatial non-uniformity. Specifically, we introduce a Spatial Signal Encoder to predict spatially adaptive modulation weights, which are used to guide multiple look-up tables for image transformation, together with an HSL-based compensation module for improved color fidelity. Beyond the architectural design, we propose an uncertainty-inspired non-uniform loss that dynamically allocates the optimization focus based on local restoration uncertainties, better matching the heterogeneous nature of real-world exposure errors. Extensive experiments demonstrate that our method achieves superior qualitative and quantitative performance compared with state-of-the-art methods. Code is available at https://github.com/FALALAS/rethinkingEC.

Figures & tables

Panel of Figure 1 — Figure 1. Motivation for Our Methodology.

Figure 2. Overview of the proposed framework. Given an incorrectly exposed input image, our method first performs non-uniform modulation estimation with a Spatial Signal Encoder to generate a spatially varying modulation signal. The estimated signal is then used to explicitly guide multiple 3D LUT bases for precise image transformation, yielding the restored output image. During training, an Uncertainty Estimator predicts a dense uncertainty map, which is used to construct the proposed uncertainty-inspired non-uniform loss $\mathcal{L}_{UNU}$ with the ground truth image. In this way, our framework jointly models spatial non-uniformity at both the architectural and optimization levels.

Figure 3. Illustration of the proposed HSL-based compensation branch. The HSL representation serves two primary purposes. It provides color-aware cues to the adaptive downsampling module and is also fed into a compensation network, where it is fused with the LUT-corrected result to produce the final refined output.

Figure 4. Visual comparisons with state-of-the-art methods on the MSEC dataset ( Afifi et al., 2021 ) .

Methods	MSEC							SICE
	Under		Over		Average			Under		Over		Average
	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
ZeroDCE (CVPR20)	14.55	0.5887	10.40	0.5142	12.06	0.5441	0.2923	16.92	0.6330	7.11	0.4292	12.02	0.5311	0.3532
RUAS (CVPR21)	13.43	0.6807	6.39	0.4655	9.20	0.5515	0.4819	16.63	0.5589	4.54	0.3196	10.59	0.4393	0.5122
SCI (CVPR22)	9.97	0.6681	5.83	0.5190	7.49	0.5786	0.3116	17.86	0.6401	4.45	0.3629	12.49	0.5051	0.4239
PairLIE (CVPR23)	11.78	0.6596	8.37	0.5887	9.73	0.6171	0.3605	16.67	0.5995	6.26	0.3846	11.47	0.4921	0.4138

Table 1. Quantitative comparisons on MSEC ( Afifi et al., 2021 ) and SICE ( Cai et al., 2018 ; Huang et al., 2022a ) . The best and second-best results are highlighted in red and blue , respectively. All results are obtained either from the officially released model weights of the compared methods or as reported in their original papers. “–” indicates that the corresponding result is unavailable due to inaccessible released code or because it was not reported in the original paper.

Figure 5. Visual comparisons with state-of-the-art methods on the SICE dataset ( Cai et al., 2018 ) .

Methods	LCDP			REED								Inference Time (s) $\downarrow$
	Average			Under		Over		Average
	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	NIQE $\downarrow$
ZeroDCE++ (TPAMI21)	18.42	0.7669	0.2204	12.58	0.6503	13.02	0.6847	12.80	0.6675	0.19	4.55	0.0024
FECNet (ECCV22)	22.34	0.8038	0.2334	16.68	0.6823	13.93	0.5881	15.31	0.6352	0.33	3.99	0.1261
LCDP (ECCV22)	23.24	0.8420	0.1368	16.75	0.7779	17.01	0.8139	16.88	0.7959	0.12	5.00	0.0472
SCI (CVPR22)	15.96	0.6646	0.2913	16.99	0.7436	15.25	0.7034	16.12	0.7235	0.14	5.07	0.0021

Table 2. Quantitative comparisons on LCDP ( Wang et al., 2022 ) and REED ( Wang et al., 2025 ) . The best and second-best results are highlighted in red and blue , respectively. All results are obtained either from the officially released model weights of the compared methods or as reported in the original papers. “–” indicates that the corresponding result is unavailable due to inaccessible released code or because it was not reported in the original paper.

Figure 6. Visual comparisons with state-of-the-art methods on the LCDP dataset ( Wang et al., 2022 ) .

Setting	Sampling		Compensation		PSNR $\uparrow$	SSIM $\uparrow$
Setting	RGB	HSL	RGB	HSL	PSNR $\uparrow$	SSIM $\uparrow$
A	bilinear				23.49	0.8732
B	✓				23.60	0.8744
C		✓			23.64	0.8751
D	✓		✓		23.62	0.8749
E		✓	✓		23.68	0.8770

Table 3. Ablation study of the color compensation design. We analyze the color spaces used in both the adaptive sampling and compensation modules. Here, "bilinear" denotes the baseline setting that employs standard bilinear downsampling.

Figure 7. Visual comparison of different modulation strategies. Settings A and B both suffer from biased color distortions, whereas Setting C (Ours) yields the most visually pleasing and accurate restoration.

Figure 8. Quantitative and structural comparison of different modulation strategies. Setting A uses plain convolution, Setting B adopts an implicit decoder, and Setting C employs explicit 3D LUT-based transformation.

Figure 9. Visualization of the predicted uncertainty maps using different inputs to the Uncertainty Estimator. As shown, the estimated uncertainty successfully distinguishes between the two distinct degradation patterns, validating its role in dynamically balancing regional optimization weight.

Method	UE Input	PSNR $\uparrow$	SSIM $\uparrow$
CoTF	–	23.44	0.8728
CoTF w/ $\mathcal{L}_{UNU}$	Raw Input	23.47	0.8737
CoTF w/ $\mathcal{L}_{UNU}$	Model Output	23.56	0.8739
Ours w/o $\mathcal{L}_{UNU}$	–	23.59	0.8762
Ours	Raw Input	23.69	0.8768
Ours	Model Output	23.72	0.8774

Table 4. Ablation study of the proposed uncertainty-inspired non-uniform loss (

\mathcal{L}_{UNU}

) and the inputs for the Uncertainty Estimator (UE).

Image Enhancement

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

See all 154 papers

Unified RAW Image Exposure Correction and Restoration via Linear Illumination and Frequency-Adaptive Enhancement
Oct 20, 2026Yingkai Zhang, Yunhao Zou, Ying FuImage Enhancement Image Restoration
School of Computer Science and Technology, Beijing Institute of Technology, Beijing, China
High-Quality Exposure Correction with Diffusion-Based Image Generation Priors
Aug 9, 2026Ziwen Li, Meng Cao, Jinpu Zhang +4Image Enhancement Diffusion Model Fine-Tuning
National Key Laboratory of Science and Technology on Multispectral Information Processing, School of Artificial Intelligence and Automation, Huazhong University of Science and Technology, Wuhan 430074, China · Mohamed bin Zayed University of Artificial Intelligence, Abu Dhabi, UAE · National University of Defense Technology, Changsha, China +1
AutoLumNet: Monotone Optimal Transport for Single-Shot Exposure Correction
Aug 20, 2026Airin Akter Tania, Md Raihan Khan, Mohiuddin AhmadImage Enhancement Low-Light Image Enhancement
Department of Electrical and Electronic Engineering (EEE) Khulna University of Engineering & Technology (KUET), Bangladesh

Oct 20, 2026J/K move · Enter open · S save

Unified RAW Image Exposure Correction and Restoration via Linear Illumination and Frequency-Adaptive Enhancement

Yingkai Zhang, Yunhao Zou, Ying Fu

School of Computer Science and Technology, Beijing Institute of Technology, Beijing, China

Image Enhancement Image Restoration

Open paper AskPDF arXiv

Incorrect exposure degrades both image brightness and recoverable scene detail. Existing sRGB-based correction methods are limited by the non-linear image signal processing pipeline, whereas RAW-based methods are usually developed for a single degradation type, such as low-light denoising or high-light recovery. In this paper, we focus on joint exposure correction and restoration directly in the RAW domain and propose a unified RAW Image Exposure Correction and Restoration (RIECR) framework for both under- and over-exposed images. RIECR decomposes the problem into two coupled sub-problems that follow the RAW imaging model: linear illumination correction and residual degradation restoration. First, a Linear Illumination Correction Network (LICN) estimates exposure-aware linear gains by using channel-specific RAW cues. It exploits the high signal-to-noise ratio of Green channels for robust brightening and the relatively unsaturated Red/Blue responses for high-light darkening. Second, a dual-branch Frequency-based Enhancement Network (FEN) restores the remaining degradation with a Frequency Dynamic Convolution (FDC) module. FDC decomposes convolutional responses into low- and high-frequency components and adaptively suppresses amplified noise or recovers clipped textures according to the exposure context. We further build a real paired RAW benchmark containing 223 scenes and 1,784 pairs covering complex multi-exposure errors. Experiments against sixteen state-of-the-art methods show that RIECR achieves state-of-the-art performance on real RAW exposure correction.

Abstract

Image Enhancement

Explore similar work

Unified RAW Image Exposure Correction and Restoration via Linear Illumination and Frequency-Adaptive Enhancement

High-Quality Exposure Correction with Diffusion-Based Image Generation Priors

AutoLumNet: Monotone Optimal Transport for Single-Shot Exposure Correction