Poor lighting can turn an ordinary photograph into a computational puzzle: details disappear into shadows, colors shift, and sensor noise becomes conspicuous when an image is brightened. A new study describes an artificial-intelligence system designed to tackle those problems in a particular order, first cleaning up the image and then separating the scene’s lighting from its underlying surface information. Called the Clean Decomposition Network, or CDNet, the method is aimed at improving photographs captured in dim environments without requiring the large collections of manually enhanced examples typically used to train supervised image-processing systems. The researchers report that CDNet performed strongly across eight public datasets, including LOL and SICE, and in some tests reached levels comparable to supervised methods while remaining an unsupervised approach.
Low-light image enhancement is more complicated than simply increasing exposure. When a camera records a dark scene, the signal arriving at each pixel is weak, so random fluctuations in the sensor and electronics can become comparable to the actual visual information. Amplifying the image increases both the useful signal and the unwanted noise. The result may be brighter but not more realistic: fine textures can be smeared away, edges can become artificial, and color relationships can be distorted. A successful enhancement system therefore has to infer which variations represent real objects and which are artifacts introduced during capture. It must also estimate how much of the darkness comes from illumination and how much reflects the properties of the objects themselves.
CDNet is based on Retinex theory, a model of image formation that treats a photograph as the product of two interacting components: illumination and reflectance. Illumination describes the light falling across the scene, including broad changes caused by shadows or weak ambient light. Reflectance represents the intrinsic appearance of surfaces—the colors, textures, and local details that would ideally remain stable if the lighting changed. In a simplified form, an observed image can be expressed as the element-by-element product of these two components. Enhancement systems using this framework attempt to estimate both components, brighten the illumination map, and preserve or reconstruct the reflectance map. The decomposition is useful because it offers a structured way to brighten a scene without uniformly washing out its details.
The difficulty is that noise is mixed into the original image before the decomposition begins. Retinex-based networks may then mistake noise for reflectance detail or interpret noisy fluctuations as changes in illumination. Either error can propagate through the rest of the enhancement process. A system that treats sensor noise as texture may produce grainy, over-sharpened photographs; one that removes genuine texture may generate unnaturally smooth surfaces. The researchers’ central strategy is to prevent this contamination from reaching the decomposition stage. CDNet first applies a denoising operation to low-light image pairs, and only afterward estimates illumination and reflectance. In principle, this ordering gives the network a cleaner representation from which to learn the difference between lighting conditions and stable scene content.
The use of image pairs is important to the network’s design. The study describes a procedure intended to establish paired low-light images whose reflectance components should remain consistent, even when their illumination differs. That consistency acts as a physical and visual constraint: a wall, face, road surface, or object should not change its underlying reflectance merely because one image is darker than another. If the network enhances one member of a pair while altering the inferred surface information, the mismatch can be used to guide its optimization. This is different from asking a model to imitate a single “correct” bright photograph. Instead, the model receives internal signals about what ought to remain unchanged, allowing it to learn from relationships within the data rather than relying exclusively on labels created by human editors.
The denoising stage is not intended simply to make the input look smoother. According to the researchers, it is designed to remove degradation while protecting details and limiting color deviation before the Retinex analysis. That balance is technically demanding because aggressive denoising can erase the very information enhancement is supposed to recover. Fine hair, fabric patterns, foliage, lettering, and object boundaries may occupy only a small number of pixels, especially in a dark image. A denoiser that removes all high-frequency variation risks confusing detail with noise. CDNet therefore combines the denoising network with a carefully designed loss function, a mathematical objective that penalizes undesirable outputs and rewards consistency among the system’s intermediate and final results. The abstract does not specify the individual terms of that loss, but it identifies the joint denoising and decomposition process as the mechanism guiding network training.
CDNet is described as unsupervised, meaning that its learning does not depend on a conventional set of paired input and target images in which every dark photograph is matched to a professionally brightened reference. Such references are difficult to produce faithfully. A human editor may choose a pleasing appearance that does not represent the scene’s actual illumination, while a photograph captured separately at a different exposure may contain changes in motion, viewpoint, or shadow. Unsupervised methods instead use constraints derived from the image itself, from relationships between related images, or from assumptions about how illumination and reflectance behave. This can make them more adaptable to real-world photographs, where the precise camera conditions and desired brightness are unknown. It also shifts the burden onto the model’s objectives: if those constraints are poorly designed, the network may satisfy them while producing visually implausible results.
The researchers evaluated the proposed method on eight publicly available datasets, with LOL and SICE specifically identified in the study. Testing across multiple datasets is intended to show whether the approach works beyond a narrow collection of scenes or camera conditions. Low-light benchmarks can differ in image content, darkness levels, noise characteristics, and the way reference or comparison images are assembled. A system that excels on one dataset may therefore struggle when the exposure pattern, sensor noise, or scene distribution changes. The authors report that CDNet achieved superior enhancement performance across the experiments and was highly competitive with existing unsupervised methods. They also report that it reached performance comparable to supervised systems on certain metrics and datasets, a notable result because supervised approaches generally benefit from direct target examples during training.
Those claims should be understood as benchmark findings rather than evidence that the network solves every low-light imaging problem. Enhancement quality can be measured in several ways, and numerical scores do not always correspond perfectly to what viewers consider natural. A metric may reward similarity to a reference image while overlooking halos, overly strong contrast, or colors that look attractive but are physically inaccurate. Conversely, a perceptually pleasing image may differ from a reference because the scene contains genuine ambiguity. The source material does not provide the study’s individual scores, dataset-by-dataset rankings, computational requirements, or comparisons with named rival models, so the scale of the reported advantage cannot be independently assessed from the abstract alone. It does establish, however, that the evaluation covered a broad set of public data and that the researchers specifically examined both unsupervised and supervised benchmarks.
If the approach proves robust outside curated datasets, its applications could extend across photography, robotics, surveillance, mobile imaging, and scientific observation. Cameras on autonomous machines may need to interpret poorly illuminated corridors or roads; phones may attempt to recover details from night scenes; and monitoring systems may operate continuously under changing light. In all of these settings, brightening without controlling noise can degrade downstream computer vision. An object detector may miss an edge hidden by grain, while a recognition system may mistake enhancement artifacts for meaningful features. By separating denoising from Retinex decomposition, CDNet reflects a broader trend in image restoration: rather than treating enhancement as a single visual adjustment, researchers are breaking the problem into stages that correspond to different sources of degradation. The study’s contribution is the argument that cleaning the signal before estimating illumination can make that separation more reliable.
The work was carried out by researchers at the Henan Institute of Science and Technology in China and was supported by several provincial and higher-education research programs in Henan. The authors declare no conflict of interest. The article was made available as an open-access, peer-reviewed accepted version with a permanent DOI, with the source noting that further edits may precede the final version of record. For now, CDNet offers a practical demonstration of how a camera’s darkest failures might be addressed: not by turning up brightness blindly, but by first asking which variations belong to the scene, which belong to the light, and which were introduced by the sensor itself. That distinction could determine whether a rescued night photograph looks merely brighter—or genuinely clear.
Subject of Research: Unsupervised low-light image enhancement using denoising and Retinex decomposition
Subject of Research: Technology and Engineering
Article Title: Clean Retinex Decomposition Network for Low-Light Image Enhancement
Article References: Ma, Y., Qiao, M., Li, X., Li, X., & Wang, H. (2026). Clean Retinex Decomposition Network for Low-Light Image Enhancement. Neural Processing Letters. https://doi.org/10.1007/s11063-026-11875-w
Image Credits: AI Generated
DOI: 10.1007/s11063-026-11875-w
Keywords: low-light image enhancement, Retinex decomposition, image denoising, Clean Decomposition Network, unsupervised learning, computational photography, image restoration
Cite Scienmag News
APA MLA Chicago
Everett Foxley. (August 28, 2026). New Retinex Network Brightens Low-Light Images Without Introducing Visual Noise. Scienmag. https://scienmag.com/new-retinex-network-brightens-low-light-images-without-introducing-visual-noise/
Everett Foxley. “New Retinex Network Brightens Low-Light Images Without Introducing Visual Noise.” Scienmag, 28 August 2026, https://scienmag.com/new-retinex-network-brightens-low-light-images-without-introducing-visual-noise/. Accessed 28 August 2026.
Everett Foxley. “New Retinex Network Brightens Low-Light Images Without Introducing Visual Noise.” Scienmag. August 28, 2026. https://scienmag.com/new-retinex-network-brightens-low-light-images-without-introducing-visual-noise/
Copy citation Download RIS
Tags: artificial intelligence for low-light photographyartificial intelligence for photographychallenges in low-light image restorationcomparison of supervised and unsupervised image enhancement methodscomputational photographydeep learning image enhancementimage decompositionimage decomposition neural networksimage noise reductionLow-light image enhancementlow-light photography challengesnoise reduction in brightened imagesnoise-free image brighteningperformance of CDNet in public datasetspreservation of image details in low-light conditionspublic dataset evaluationRetinex networkRetinex-based image processingscene illumination separationsmart low-light photo enhancement techniquesunsupervised deep learning for image denoisingunsupervised image processing
