RetinexGuI: Retinex-Guided Iterative Illumination Estimation Method for Low Light Images

Yasin Demir, Nur Hüseyin Kaplan, Sefa Kucuk, Nagihan Severoglu · 2026

Read RetinexGuI: Retinex-Guided Iterative Illumination Estimation Method for Low Light Images by Yasin Demir, Nur Hüseyin Kaplan, Sefa Kucuk, Nagihan Severoglu (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-13320.txt.

Introduction

Yasin Demir, Nur Huseyin Kaplan, Sefa Kucuk, and Nagihan Severoglu brighten low light pics with a guide from a classic rule. The Retinex rule splits a pic into light fall and true surface shades. Bright pics from dark ones aid key tasks that run at night. Both rule based and learnt tools do well at the task.

Yet many tools are heavy or need huge train sets. That caps their use in live apps. The team's method, RetinexGuI, steps the light guess in rounds, led by the Retinex split. Each round lifts the guess with less cost. For cam and phone teams, a light tool that runs live has clear worth.

Catalog record

CREATORYasin Demir, Nur Hüseyin Kaplan, Sefa Kucuk, Nagihan Severoglu
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE14 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-13320

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

RetinexGuI: Retinex-Guided Iterative Illumination Estimation Method for Low Light Images Yasin Demir, Nur Hüseyin Kaplan, Sefa Kucuk, Nagihan Severoglu Abstract In recent years, there has been a growing interest in low-light image enhancement (LLIE) due to its importance for critical downstream tasks. Current Retinex-based methods and learning-based approaches have shown significant LLIE performance. However, computational complexity and dependencies on large training datasets often limit their applicability in real-time applications. We introduce RetinexGuI, a novel and effective Retinex-guided LLIE framework to overcome these limitations. The proposed method first separates the input image into illumination and reflection layers, and iteratively refines the illumination while keeping the reflectance component unchanged. With its simplified formulation and computational complexity of 𝒪 ⁡ ( N ) \displaystyle\mathcal{O}(N) , our RetinexGuI demonstrates impressive enhancement performance across three public datasets, indicating strong potential for large-scale applications. Furthermore, it opens promising directions for theoretical analysis and integration with deep learning approaches. The source code will be made publicly available at https://github.com/etuspars/RetinexGuI once the paper is accepted. I Introduction As real-time imaging systems become more common, capturing reliable images in low-light conditions is increasingly important. However, these images often suffer from reduced visibility and amplified noise, which can degrade downstream vision tasks. Therefore, the development and implementation of fast and efficient low-light image enhancement (LLIE) methods are crucial for improving visual quality and supporting practical real-time applications. Currently, LLIE research is largely dominated by two main approaches: Retinex-based methods [ 4 , 5 , 16 , 31 , 18 , 26 , 23 , 35 , 6 , 3 ] and learning-based methods [ 33 , 20 , 10 , 14 , 22 , 2 , 21 , 17 , 15 , 27 , 24 , 32 , 30 , 34 , 7 ] . Retinex-based methods model an observed image as the product of illumination and reflectance components under the Retinex assumption [ 13 ] . Early versions, including single-scale Retinex [ 12 ] and multi-scale Retinex [ 11 ] , treat the estimated reflectance as the final enhanced output. However, the results produced by these methods are often prone to over-enhancement and color distortion, leading to visually unnatural appearances. Recent Retinex-based approaches formulate LLIE as a variational optimization problem with explicit regularization on both components [ 31 , 35 , 6 , 3 ] . These variational Retinex models typically rely on hand-crafted regularization terms and require careful parameter tuning, which increases computational cost and limits their real-time applicability. Learning-based algorithms, on the other hand, leverage deep neural networks to learn complex enhancement mappings from data, covering both learning-driven [ 22 , 10 , 14 , 17 ] and Retinex-inspired approaches [ 20 , 2 , 34 ] . Although they can produce impressive visual results, such models typically require large training datasets, substantial computational resources, and task-specific training, which limit their flexibility and suitability for real-time applications. To overcome the aforementioned limitations, we introduce a novel Retinex-guided LLIE framework that provides fast and stable enhancement with minimal computational cost. In contrast to prior Retinex-based methods, we present a lightweight and computationally efficient enhancement strategy that iteratively refines the illumination component while keeping the reflectance unchanged, without relying on any external prior or complex regularization. We can list the key contributions of the RetinexGuI as follows: • A simplified formulation: Compared to recent Retinex-based approaches, our framework employs a simplified Retinex-guided formulation and does not require external priors or hyperparameter tuning. • Computational efficiency: The proposed method eliminates the need for extensive training datasets or heavy computational resources. This provides a good balance between efficiency and visual quality, supporting real-time performance for various low-light enhancement tasks. • Robust and versatile enhancement behavior: Our RetinexGuI performs reliably in various low-light scenarios and offers potential for extension to related applications, such as virtual exposure generation. II Proposed RetinexGuI Method Based on Retinex theory [ 13 ] , an observed image \displaystyle\mathbf{I} is expressed in terms of its illumination and reflectance components as \mathbf{I}=\mathbf{L}\odot\mathbf{R} (1) here, \displaystyle\mathbf{L} and \displaystyle\mathbf{R} represent the illumination and reflectance components, respectively, and \displaystyle\odot indicates element-wise multiplication. For convenience, we reformulate the Retinex model in the logarithmic domain as \displaystyle\mathbf{I}_{l}=\mathbf{L}_{l}+\mathbf{R}_{l} (2) where \displaystyle\mathbf{I}_{l}=\log(\mathbf{I}) , \displaystyle\mathbf{L}_{l}=\log(\mathbf{L}) , and \displaystyle\mathbf{R}_{l}=\log(\mathbf{R}) . We iteratively enhance the input image by updating the illumination component while keeping the reflectance component unchanged. Specifically, the enhanced image in the logarithmic domain is obtained as \mathbf{I}_{l}^{(N+1)}=\mathbf{L}_{l}^{(N)}+\mathbf{R}_{l}, (3) where \displaystyle N denotes the iteration number. The initial values are set as \displaystyle\mathbf{I}^{(0)}=\mathbf{I} and \displaystyle\mathbf{R}_{l}=\mathbf{I}^{(0)} . The illumination component is then iteratively estimated by dividing the observed image by the initial reflectance component as \mathbf{L}_{l}^{(N)}=\frac{\mathbf{I}_{l}^{(N)}}{\mathbf{R}_{l}}. (4) Given that \displaystyle\mathbf{R}_{l}=\mathbf{I}^{(0)} , we can rewrite ( 4 ) as \mathbf{L}_{l}^{(N)}=\dfrac{\mathbf{I}_{l}^{(N)}}{\mathbf{I}_{l}^{(0)}}. (5) By substituting ( 5 ) into ( 3 ), the update equation can be equivalently expressed as \mathbf{I}_{l}^{(N+1)}=\dfrac{\mathbf{I}_{l}^{(N)}}{\mathbf{I}^{(0)}_{l}}+\mathbf{I}^{(0)}_{l}. (6) The recursive update given in ( 6 ) is equal to the following: \mathbf{I}_{l}^{(N+1)}=\frac{\sum_{n=0}^{N}\left(\mathbf{I}_{l}^{(0)}\right)^{n}}{\left(\mathbf{I}_{l}^{(0)}\right)^{(N)}}+\mathbf{I}_{l}^{(0)}=\sum_{n=0}^{N}\left(\mathbf{I}_{l}^{(0)}\right)^{-n}+\mathbf{I}_{l}^{(0)}. (7) When the number of iterations becomes sufficiently large ( \displaystyle N\to\infty ), the formulation converges as follows: \mathbf{I}_{l}^{(\infty)}=\frac{1}{1-\left(\mathbf{I}_{l}^{(0)}\right)^{-1}}+\mathbf{I}_{l}^{(0)}=\frac{\left(\mathbf{I}_{l}^{(0)}\right)^{2}}{\mathbf{I}_{l}^{(0)}-1}. (8) It should be noted that all operations are performed on an element-wise basis. The closed-form result demonstrates that iterative illumination refinement converges to a solution defined by the initial illumination estimate. As the number of iterations increases, the enhancement process becomes independent of intermediate updates and is managed by a simple analytical formulation. This property explains the stable behavior of the proposed RetinexGuI and its ability to achieve consistent enhancement results with only a few iterations. The result of ( 8 ) is considered as the initial enhancement level. To obtain further levels of enhancement, especially for extremely dark scenes, ( 8 ) is applied in a cascaded manner. III Experiments and Results We perform all experiments on a PC equipped with Windows 11 OS, Intel Core i7-11800H CPU at 2.3 GHz, 16 GB of RAM, Nvidia RTX 3060 6 GB GPU, and MATLAB2024b . We compare the RetinexGuI with five related methods, including SRIE [ 4 ] , Hu \displaystyle et~al. [ 6 ] , Wang \displaystyle et~al. [ 28 ] , GCP-MCS [ 9 ] , and ACGC [ 25 ] . For a fair comparison, we obtain the implementation codes from the authors’ official websites and follow the recommended parameter settings. We evaluate the performance of competing methods on three well-known low-light datasets: VE-LOL-L [ 19 ] , RELLISUR [ 1 ] , and LoLi-Street [ 8 ] . Among them, the VE-LOL-L dataset consists of 1000 synthetic paired and 500 real paired images; the RELLISUR dataset includes 85 paired images, each with five different exposure levels; the LoLi-Street dataset contains 3k paired images. III-A Implementation Details We first transform the input RGB-color low-light image into the Hue-Saturation-Value (HSV) space. Since the V channel represents illumination and reflectance information, we perform iterative illumination estimation on the V component. To handle under-exposed or extra-dark images, we repeatedly apply the converged illumination expression in ( 8 ). We categorize the input images into three groups based on the mean value \displaystyle(\mu_{V}) of the V component. The number of levels \displaystyle(K) is determined as \displaystyle K=\begin{cases}1,&\mu_{V}>0.16,\\ 2,&0.08\leq\mu_{V}\leq 0.16,\\ 3,&\mu_{V}<0.08.\end{cases} (9) The converged illumination update in ( 8 ) is then applied iteratively as \displaystyle\mathbf{V}_{l}^{k+1} \displaystyle=\frac{\left(\mathbf{V}_{l}^{k}\right)^{2}}{\mathbf{V}_{l}^{k}-1},\qquad

Read the complete text