Source-linked AI summary
Low-light Image Enhancement via Breaking Down the Darkness
Qiming Hu, Xiaojie Guo
TL;DR
Low-light enhancement must handle dim illumination together with amplified noise and color distortion. Bread divides these degradations in YCbCr, using specialized processing for luminance and chrominance. The paper reports qualitative and quantitative advantages over state-of-the-art alternatives across benchmark datasets.
Problem
Low-light images contain entangled dim light, noise, and color distortion, while illumination adjustment can amplify hidden, spatially correlated noise.
Method
Bread converts RGB images to YCbCr and uses illumination adjustment, illumination-guided noise suppression in luminance, and luminance-guided chrominance adaptation.
Results
Bread outperforms state-of-the-art alternatives qualitatively and quantitatively across several benchmark datasets.
Takeaways & Limitations
Separating texture and color enables specialized treatment of noise and color distortion alongside light adjustment within low-light enhancement.
Abstract
from arXiv · showhide
Images captured in low-light environment often suffer from complex degradation. Simply adjusting light would inevitably result in burst of hidden noise and color distortion. To seek results with satisfied lighting, cleanliness, and realism from degraded inputs, this paper presents a novel framework inspired by the divide-and-rule principle, greatly alleviating the degradation entanglement. Assuming that an image can be decomposed into texture (with possible noise) and color components, one can specifically execute noise removal and color correction along with light adjustment. Towards this purpose, we propose to convert an image from the RGB space into a luminance-chrominance one. An adjustable noise suppression network is designed to eliminate noise in the brightened luminance, having the illumination map estimated to indicate noise boosting levels. The enhanced luminance further serves as guidance for the chrominance mapper to generate realistic colors. Extensive experiments are conducted to reveal the effectiveness of our design, and demonstrate its superiority over state-of-the-art alternatives both quantitatively and qualitatively on several benchmark datasets. Our code is publicly available at https://github.com/mingcv/Bread.
1. Introduction
Low-light enhancement must address entangled dim light, noise, and color distortion rather than adjust illumination alone. Bread separates texture and color in YCbCr so denoising, illumination adjustment, and color correction can be handled specifically.
- Low-light photographs suffer poor visibility, while longer exposure can introduce blur and flash can create unwanted highlights and unbalanced lighting.
- Illumination adjustment amplifies hidden noise, while insufficient photons can also produce color distortion even after lighting and noise are handled.
- Bread converts RGB images into luminance and chrominance components, assigning noise removal and light adjustment to texture and color correction to the other component.
- The YCbCr space is selected as a candidate for texture-color decomposition after visualizing noise distributions and single-channel restoration effects.
- The framework introduces illumination-guided noise synthesis for amplified, spatially correlated luminance noise and a color adaption network guided by luminance.
- Extensive comparisons and ablation studies evaluate Bread qualitatively and quantitatively against state-of-the-art methods.
2. Related Work
Prior low-light enhancement methods include traditional intensity and Retinex techniques alongside deep learning approaches. Existing methods improve aspects of enhancement but leave complex, multientangled degradation insufficiently addressed.
- Traditional Methods: Traditional methods adjust value ranges, apply Gamma correction, or expand dynamic range with global and local histograms.
- Traditional Methods: Traditional enhancement quality is difficult to guarantee because these methods are content-blind.
- Traditional Methods: Retinex-based methods estimate illumination using Gaussian blurs or local maxima, then manipulate illumination to produce enhanced results.
- Deep Learning-based Methods: Deep learning methods use mechanisms such as MSR, BM3D denoising, synthetic Gamma-and-noise pairs, and auto-encoder mappings.
- Deep Learning-based Methods: Recent approaches can reduce paired-data requirements or improve enhancement, but often focus mainly on lighting and remain insufficient for complex multientangled degradation.
3. Methodology
Bread separates low-light degradation into luminance-related illumination and noise versus chrominance-related color, then addresses these factors with specialized networks. It converts RGB to YCbCr, adjusts illumination, suppresses amplified noise adaptively, and uses enhanced luminance to guide color adaptation.
- Problem Analysis & Motivation: Bread converts RGB images into YCbCr to separate luminance and chrominance, enabling specialized handling of illumination, noise, and color degradation.The framework follows a divide-and-rule design motivated by the entanglement of dim light, amplified noise, and color distortion.
- ANSN: Adaptive Noise Suppression Network: The adaptive noise suppression network uses illumination-guided noise levels to remove noise amplified during luminance adjustment.The method synthesizes amplified Gaussian noise on normal-light references and conditions suppression on an attention map derived from relative illumination.
- Method Architecture: Bread comprises illumination adjustment, adaptive noise suppression, and color adaption networks, with uniform three-downsampling and three-upsampling sub-network structures.The sub-networks are optimized individually and use different tail activations, with ANSN remaining convolutional at its output.
- IAN: Illumination Adjustment Network: The illumination adjustment network predicts a relative illumination map and uses it to transform low-light luminance toward normal-light luminance.Its objective includes illumination estimation, piece-wise smoothness, and gradient-domain similarity to reduce halo artifacts.
- ANSN: Adaptive Noise Suppression Network: A noise fusion module combines luminance outputs produced with different suppression strengths to improve robustness against over-smoothing and preceding-stage errors.The fused luminance is learned against the normal-light target using reconstruction and structural-similarity objectives.
- CAN: Color Adaption Network: The color adaption network maps chrominance using enhanced luminance guidance, training with reference luminance and testing with fused enhanced luminance.Multi-exposure data is introduced to address low-saturation training references and encourage colors appropriate to a real exposure condition.
4. Experiments
Experiments evaluate Bread on multiple low-light datasets using quantitative metrics, visual comparisons, and ablations. The framework achieves strong results, with noise suppression, reduced artifacts, and improved color realism reported qualitatively.
- Bread is evaluated on LOL, DICM, NPE, and VV using PSNR, SSIM, NIQE, DeltaE, and LOE.
- Bread outperforms state-of-the-art alternatives by a noticeable margin on both reference and no-reference metrics on the LOL evaluation dataset.The comparison is reported in Table 1.
- Visual results show that Bread produces higher-quality images with well-suppressed noise and fewer artifacts than competing methods.Other methods exhibit irregular illumination, noise residuals, and texture or color distortion.
- Multi-exposure-data-based color adaptation rectifies greenish hue and produces more realistic tones.Figure 3 also reports more vibrant color when multi-exposure data are introduced.
- Bread occupies the leading position on both non-reference metrics across three benchmark datasets when compared with data-driven methods trained on LOL.The paper cautions that NIQE does not monotonically track visual noise.
- Ablations assess denoising, noise fusion, IA-NS separation, noise synthesis models, and multi-exposure data to evaluate framework configurations.The study compares fixed Gaussian, Poisson, and spatially-correlated Gaussian noise synthesis.
- Removing IA-NS separation creates a performance gap relative to noise-synthesis approaches, emphasizing the separation strategy’s importance.
5. Conclusion
The conclusion presents Bread as a divide-and-rule framework that disentangles low-light degradation through luminance-chrominance processing. Experiments validate its subnetwork-specific design, and the principle may extend to other enhancement and restoration tasks.
- Bread converts RGB images to YCbCr so luminance handles brightening and denoising while chrominance handles color correction guided by enhanced luminance.IAN, ANSN, and CAN are customized and trained individually for specific degradation types.
- Experiments validate that separating degradation-specific subnetworks creates a training environment tailored to individual restoration problems.
- The divide-and-rule texture-color decomposition principle may also apply to dehazing and underwater image enhancement.