Source-linked AI summary
Retinexmamba: Retinex-based Mamba for Low-light Image Enhancement
Jiesong Bai, Yuhao Yin, Qiyuan He, Yuanxian Li, Xiaofeng Zhang
TL;DR
Low-light enhancement methods face difficulties preserving details, controlling noise and artifacts, and modeling image-wide dependencies while retaining physical interpretability. RetinexMamba combines Retinex-inspired illumination estimation with an SS2D-based damage restorer and replaces IG-MSA with IFA. On the LOL dataset, it outperforms prior deep-learning methods based on Retinex theory in quantitative and qualitative evaluations, although its overall parameter count increases.
Problem
Low-light enhancement must address noise, artifacts, detail preservation, computational complexity, and limited interpretability in existing Retinex-based and attention-based methods.
Method
RetinexMamba combines an illumination estimator with a damage restorer using SS2D and replaces Retinexformer's IG-MSA with Illumination Fusion Attention.
Results
RetinexMamba outperforms current state-of-the-art methods on the LOL dataset in reported quantitative and qualitative evaluations.
Takeaways & Limitations
RetinexMamba provides a Retinex-based low-light enhancement architecture that combines illumination modeling, artifact restoration, and Mamba-based processing.
Takeaways & Limitations
Although SS2D reduces computational complexity, RetinexMamba has an increased overall parameter count and consumes more computational resources.
Abstract
from arXiv · showhide
In the field of low-light image enhancement, both traditional Retinex methods and advanced deep learning techniques such as Retinexformer have shown distinct advantages and limitations. Traditional Retinex methods, designed to mimic the human eye's perception of brightness and color, decompose images into illumination and reflection components but struggle with noise management and detail preservation under low light conditions. Retinexformer enhances illumination estimation through traditional self-attention mechanisms, but faces challenges with insufficient interpretability and suboptimal enhancement effects. To overcome these limitations, this paper introduces the RetinexMamba architecture. RetinexMamba not only captures the physical intuitiveness of traditional Retinex methods but also integrates the deep learning framework of Retinexformer, leveraging the computational efficiency of State Space Models (SSMs) to enhance processing speed. This architecture features innovative illumination estimators and damage restorer mechanisms that maintain image quality during enhancement. Moreover, RetinexMamba replaces the IG-MSA (Illumination-Guided Multi-Head Attention) in Retinexformer with a Fused-Attention mechanism, improving the model's interpretability. Experimental evaluations on the LOL dataset show that RetinexMamba outperforms existing deep learning approaches based on Retinex theory in both quantitative and qualitative metrics, confirming its effectiveness and superiority in enhancing low-light images.
1 Introduction
RetinexMamba combines Retinex-inspired illumination estimation with Mamba-based restoration to address low-light artifacts, computational complexity, and attention interpretability. Experiments report superiority over prior deep-learning methods based on Retinex theory.
- Low-light images commonly suffer from dimness, blurriness, and unclear details, reducing their quality and usability in vision applications.
- Retinex theory separates illumination and reflection conceptually, but traditional enhancement methods struggle with complex lighting dynamics and natural image preservation.
- RetinexMamba divides enhancement into an illumination estimator and damage repairer, initially brightening images before removing amplified artifacts and noise.The repairer also addresses color distortions and overexposure.
- SS2D replaces Transformer processing to capture long-range dependencies with linear computational efficiency, while IFA replaces IG-MSA to improve attention interpretability.
- Comprehensive quantitative and qualitative experiments on the LOL dataset report that RetinexMamba surpasses state-of-the-art deep-learning methods based on Retinex theory.
- RetinexMamba introduces Mamba for low-light enhancement and uses illumination-feature fusion designed to remain consistent with Retinex theory.
2 Related Work
Earlier low-light enhancement methods used distribution mapping, Retinex decomposition, CNNs, and emerging state space models. Their reported challenges include artifacts, parameter-sensitive illumination estimation, noise amplification, and limited long-range dependency modeling.
- Distribution-mapping methods such as histogram equalization and gamma correction can produce color distortion and artifacts because they do not use semantic information.
- Retinex methods explain enhancement through illumination removal but often depend on handcrafted priors and parameter tuning, with inaccurate estimates causing artifacts, color bias, and noise retention or amplification.
- CNN-based enhancement methods capture local spatial information but have limitations in modeling long-range dependencies across image regions.
- State space models emerged as alternatives to CNNs and Transformers for modeling long-range dependencies, with Mamba among the contemporary developments in this direction.
3 Method
RetinexMamba combines Retinex-inspired illumination estimation with an IFVM damage restorer, using illumination-guided attention and SS2D-based state-space processing to enhance low-light images while addressing artifacts and computational cost.
- Architecture: RetinexMamba combines an illumination estimator with an IFVM-based damage restorer built around IFSSM.The estimator produces an illuminated image and illumination features that feed the damage restorer.
- Retinex-based Framework: Perturbation modeling extends Retinex decomposition with illumination and reflection perturbations to represent quality losses amplified during enhancement.The formulation adds perturbation terms to the conventional element-wise illumination–reflection model.
- Illumination Estimator: The illumination estimator combines the low-light image with an illumination prior, then uses convolutions to generate an illuminated image and illumination feature map.The prior is computed from image-channel averages and supplies lighting information for enhancement.
- Damage Restorer: The damage restorer uses a symmetric two-level encoder–decoder with downsampling, upsampling, and IFSSM blocks to reconstruct enhanced images.Its components include Layer Normalization, IFA, SS2D, FFN, and convolutional dimension matching.
- Illumination-Fused Attention: IFSSM replaces Retinexformer's IG-MSA with Illumination Fusion Attention, using illumination features as queries and image features as keys and values.This design fuses lighting information with the input while preserving query-key-value consistency within the input representation.
2D-Selective Scan.
The SS2D module converts image features into directional sequences, processes them with S6 blocks, and merges the results back into image format for efficient visual feature extraction.
- 2D-Selective Scan: SS2D scans images from four directions, flattens spatial dimensions into token sequences, processes them with S6 blocks, and merges the directional outputs.The merged representation is reshaped to match the original input size.
- 2D-Selective Scan: SS2D increases its hidden-state depth across sampling levels, using dstate 16 by default and up to 64 at the deepest level.The hidden-layer count doubles with each sampling level.
4 Experiment
RetinexMamba was evaluated on LOL datasets against state-of-the-art methods using quantitative and qualitative comparisons, plus ablations of its design components. It achieved stronger PSNR and RMSE results overall, while qualitative results showed reduced noise, color distortion, and exposure errors.
- Datasets and Implementation details: The evaluation used LOLv1, LOLv2_real, and LOLv2_synthetic datasets, with LOLv1 containing 485:15 training-to-test image pairs.The model was evaluated on both versions of the LOL benchmark, while the ablation study covered three dataset variants.
- Quantitative Results: RetinexMamba outperformed compared state-of-the-art methods in PSNR and RMSE, although its SSIM was slightly lower than Retinexformer.Comparisons covered synthetic LOLv1 data and real LOLv2 data under matched conditions.
- Quantitative Results: 0.093 and 0.77 PSNR gains were achieved on LOL_v1 and LOLv2_real, respectively, while RMSE decreased by 0.39 on LOLv2_real.These improvements were reported against the baseline and best existing technology method; lower RMSE is better.
- Qualitative Results: Qualitative comparisons showed that RetinexMamba controlled exposure intensity, reduced color distortion, minimized noise, and produced HSV results closest to the Ground Truth.Other methods exhibited noise amplification, underexposure, overexposure, or color distortion in the compared images.
- Ablation Study: The full ablation setup achieved the highest PSNR and SSIM among variants that fixed hidden-layer depth, removed fused blocks or SS2D, or replaced fused attention with IG-MSA.The variants tested feature extraction depth, illumination fusion, SS2D, and attention design.
5 Conclusion
The conclusion presents RetinexMamba as a Retinexformer- and Mamba-based architecture for low-light enhancement. It reports state-of-the-art performance on LOL, while noting that increased parameters remain a computational-resource limitation.
- 5 Conclusion: RetinexMamba combines Retinexformer-style illumination estimation and damage restoration with Mamba-based SS2D and fused attention for low-light enhancement.The architecture replaces IG-MSA with a more interpretable Fused-Attention module and uses SS2D for visual feature processing.
- 5 Conclusion: Extensive quantitative and qualitative experiments demonstrate that RetinexMamba outperforms current state-of-the-art methods on the LOL dataset.
- 5 Conclusion: Although SS2D reduces computational complexity, the overall parameter count increases and consumes more computational resources.The paper identifies parameter reduction as future work while maintaining computational complexity.