Source-linked AI summary
IDM-Net: A Lightweight Illumination-Decoupled Modulation Network for Low-Light Image Enhancement
Cheng-Yen Hsiao, Jing-Ming Guo
TL;DR
Lightweight low-light enhancement must balance brightness restoration with color fidelity, while existing color-decoupled methods leave luminance’s role as a guidance prior underexplored. IDM-Net uses luminance priors to guide RGB feature reconstruction and reports competitive performance among lightweight methods, though fixed priors may constrain generalization in some scenarios.
Problem
RGB low-light enhancement couples luminance and chrominance, and recent color-decoupled methods primarily optimize luminance rather than exploiting it as a feature-reconstruction guidance prior.
Source: §1 · pp. 1–2Method
IDM-Net uses a dual encoder to extract RGB structure features and luminance priors, which IGM injects into reconstruction, alongside an FRB for detail recovery.
Source: §1 · p. 2, §3.0.1 · p. 5, §3.0.3 · p. 7Results
24.15 dB PSNR and 0.845 SSIM on LOL-v1; IDM-Net achieves the best overall performance among lightweight methods.
Source: §4.0.3 · p. 9Takeaways & Limitations
IDM-Net’s feature-modulation approach combines multi-scale luminance guidance with structural features to restore brightness while preserving structural and chromatic fidelity.
Source: §5 · p. 11Photometric inconsistencies and a fixed illumination prior may limit generalization to some real-world camera imaging pipelines.
Source: §5 · p. 12
Abstract
from arXiv · showhide
Low-light image enhancement (LLIE) remains challenging for lightweight models because illumination restoration and color fidelity are difficult to optimize simultaneously in the RGB color space. Although recent color-decoupled methods separate luminance and chrominance representations, they primarily optimize luminance as an enhancement target, leaving its potential as an explicit guidance prior largely unexplored during feature reconstruction. To address this limitation, we propose IDM-Net, a lightweight Illumination-Decoupled Modulation Network for low-light image enhancement. IDM-Net adopts a dual-encoder architecture consisting of a structure encoder that extracts multi-scale appearance features from the RGB image and a lightweight illumination encoder that learns illumination priors from the decoupled luminance (Y) channel. To effectively exploit these priors, we introduce an Illumination-Guided Modulation (IGM) module that injects multi-scale illumination cues into the decoder through spatially adaptive affine modulation, enabling accurate brightness restoration while preserving natural color consistency. Furthermore, we design a lightweight Feature Refinement Block (FRB) to progressively suppress degradation artifacts and recover fine-grained image details during reconstruction. Extensive experiments on multiple standard low-light image enhancement benchmarks demonstrate that IDM-Net achieves competitive performance among lightweight LLIE methods while maintaining an excellent balance between restoration quality and computational efficiency.
1 Introduction
Lightweight low-light enhancement must balance brightness restoration with color fidelity, but RGB methods couple luminance and chrominance. IDM-Net uses luminance as an illumination prior to guide RGB reconstruction, and combines a dual encoder, IGM, and FRB for competitive quality and efficiency.
1 Introduction
- RGB-based LLIE methods struggle to optimize brightness restoration and color fidelity together because luminance and chrominance are entangled.Low-light images can also suffer low brightness, severe noise, and color distortion, affecting visual quality and downstream vision tasks.Source: §1 · pp. 1–2
- Recent color-decoupled methods separate illumination and color, but primarily treat luminance as an enhancement target rather than a guidance prior for feature reconstruction.The paper identifies this gap as limiting the use of illumination cues to guide restoration under complex lighting.Source: §1 · p. 2
- IDM-Net uses a dual-encoder architecture and Illumination-Guided Modulation to exploit luminance as an explicit prior for feature reconstruction.It also introduces a lightweight Feature Refinement Block for detail recovery; experiments report competitive performance and a quality-efficiency trade-off.Source: §1 · p. 2
2 Related Works
Related work spans diverse LLIE approaches, from learned low-to-normal mappings to generative and attention-based models. It also covers illumination and color decoupling, and lightweight designs; IDM-Net is motivated by limited integration of illumination cues into feature reconstruction and the efficiency-quality trade-off.
2.0.1 Low-light Image Enhancement
- LLIE work ranges from CNN mappings and task-decoupled designs to generative, transformer, state-space, and diffusion approaches.Examples include unsupervised adversarial enhancement, normalizing flows, illumination-aware attention, linear-complexity state-space modeling, and iterative denoising.Source: §2.0.1 · pp. 2–3
2.0.2 Illumination Modeling and Color-Decoupled Enhancement
- Illumination modeling and alternative color spaces seek brightness restoration with color fidelity, but prior methods provide limited interaction between illumination cues and feature reconstruction.IDM-Net addresses this gap with an illumination encoder and IGM to inject multi-scale priors into RGB feature decoding.Source: §2.0.2 · p. 3
2.0.3 Lightweight LLIE Networks
- Lightweight LLIE methods use varied priors and efficient architectures, yet face a quality-efficiency trade-off and often insufficiently integrate illumination into feature reconstruction.IDM-Net combines illumination-guided modulation with lightweight feature refinement; Figure 1 describes luminance features as priors guiding multi-scale RGB reconstruction rather than as a direct restoration target.Source: §2.0.3 · pp. 3–4, Figure 1 · p. 4
3 Methodology
IDM-Net separates appearance and illumination encoding, then uses illumination-guided modulation and feature refinement during decoding. Its design pairs multi-scale structural features with luminance-derived priors for reconstruction.
3.0.1 The Illumination-Decoupled Dual Encoder
- RGB input and BT.601 luminance feed parallel encoders that produce complementary multi-scale appearance and illumination features.The structural encoder preserves structural and chromatic information, while the illumination encoder learns illumination-aware representations from Y.Source: §3.0.1 · p. 5
3.0.2 Illumination-Guided Modulation (IGM)
- IGM injects illumination priors through spatially adaptive affine modulation, guiding structural reconstruction at each decoder scale while preserving chromatic consistency.Spatial attention emphasizes illumination-sensitive regions, and dilated context modeling captures broader lighting distributions.Source: §3.0.2 · p. 6
3.0.3 Feature Refinement Block (FRB)
- FRB combines multi-scale dilated-convolution attention to suppress noise while preserving coherent structure with gated feature modulation for selective refinement.MSAE localizes noise across scales; GFM selectively amplifies informative responses and suppresses redundant ones.Source: §3.0.3 · pp. 6–7
4 Experimental Results
IDM-Net is evaluated on paired and unpaired low-light benchmarks for restoration quality and efficiency. Its results show strong performance among lightweight methods, while ablations assess the contributions of its components.
4.0.1 Datasets and Evaluation Metrics
- Evaluation uses LOLv1, LOLv2-real, and LOLv2-synthetic with PSNR, SSIM, and LPIPS, plus five unpaired datasets assessed by NIQE.The unpaired datasets are LIME, MEF, NPE, VV, and DICM; they lack ground-truth references.Source: §4.0.1 · p. 8
4.0.2 Implementation Details
- The model uses 0.79M parameters and 5.47 GFLOPs in the implementation setup, with three Feature Refinement Blocks stacked at each decoder scale unless specified otherwise.Training uses original-resolution images, random flipping and rotation, and 800 epochs with AdamW and cosine annealing.Source: §4.0.2 · p. 8
4.0.3 Quantitative Results
- Among lightweight methods, IDM-Net leads overall, reaching 24.15 dB PSNR and 0.845 SSIM on LOL-v1 and 25.11 dB PSNR and 0.935 SSIM on LOL-v2-Synthetic.On LOL-v2-Real, it achieves 23.01 dB PSNR and 0.858 SSIM, ranking among the top lightweight methods.Source: §4.0.3 · p. 9, Table 1 · p. 9
- IDM-Net achieves the best average NIQE score, 3.191, among lightweight methods on five unpaired datasets, ranking best on NPE, VV, and DICM and second-best on MEF and LIME.Lower NIQE indicates better perceptual image quality.Source: §4.0.3 · pp. 9–10, Table 2 · p. 10
- On representative unpaired images, IDM-Net generally produces more natural results with balanced illumination, realistic color reproduction, and preserved details.The qualitative comparison covers methods on representative images from unpaired low-light datasets.Source: §4.0.3 · p. 10, Figure 4 · p. 11
4.0.4 Ablation Study
- On LOL-v1, replacing standard blocks with LightBottleneck raises PSNR from 21.45 to 22.06 dB and SSIM from 0.798 to 0.812; adding illumination features by concatenation reaches 23.26 dB PSNR and 0.836 SSIM.The baseline LPIPS is 0.199, and LightBottleneck reduces it to 0.169; Table 3 summarizes the component ablation.Source: §4.0.4 · p. 10, Table 3 · p. 10
- Replacing concatenation with IGM improves PSNR from 23.26 to 23.86 dB, SSIM from 0.836 to 0.840, and LPIPS from 0.133 to 0.130; adding FRB yields the full model.The full model reaches 24.15 dB PSNR, 0.845 SSIM, and 0.129 LPIPS, with baseline gains of 2.70 dB PSNR and 0.047 SSIM and a 0.07 LPIPS reduction.Source: §4.0.4 · pp. 10–11, Table 3 · p. 10
5 Conclusion, Limitation and Future Works
IDM-Net uses luminance priors to modulate structural features, with FRB supporting reconstruction. It is competitive among lightweight LLIE methods, though photometric inconsistencies and fixed illumination priors may constrain generalization.
5 Conclusion, Limitation and Future Works
- IDM-Net uses a dedicated illumination encoder and IGM to guide structure features with multi-scale luminance priors, alongside FRB for reconstruction.The conclusion describes this decoupling as supporting brightness restoration while preserving structural and chromatic fidelity.Source: §5 · p. 11
- Benchmark results show highly competitive performance among lightweight LLIE methods.Source: §5 · p. 11
- Photometric inconsistencies and a fixed illumination prior may limit generalization to some scenarios with diverse camera imaging pipelines.Source: §5 · p. 12
- Future work will investigate learnable illumination priors and an ultra-lightweight version for real-time deployment on resource-constrained devices.Source: §5 · p. 12