Low-light Image Enhancement via Multi-scale Attention combined with Fourier Transform

arXiv:2607.24002 · cs.CV · Submitted 2026-07-27 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "Low-light Image Enhancement via Multi-scale Attention combined with Fourier Transform".

Jane: Here's the gist of the paper. Low-light image enhancement (LLIE) aims to improve image quality and clarity in diverse and demanding low-illumination environments.

Tom: This paper presents a supervised frequency domain deep learning network named multi-scale attention combined with the Fourier transform (MSFT) for LLIE.

Jane: The proposed MSFT method is a two-stage U-shaped framework combining the Fourier transform with a CNN-Transformer Cross-mixed U-net framework.

Tom: This architecture involves several key components including the Retinex decomposition performed in the original image space.

Title and authors: Tom: Let's look at what they are actually trying to achieve with this "Low-light Image Enhancement via Multi-scale Attention combined with Fourier Transform" framework. It’s not just one trick; it’s a whole system designed to solve the problem of low-light image enhancement, which is often messy.

Jane: The summary points out that existing deep learning methods struggle to capture real-world illumination and restore texture details because their algorithmic strengths aren't being fully used. This paper directly addresses that by proposing a supervised frequency domain deep learning network named MSFT.

Lu: They detail the U-shaped framework, which combines a CNN-Transformer cross-mixed U-net architecture, and they use Retinex decomposition in the original image space as part of that process to get started.

Meng: The core enhancement module they focus on is called Fourier Transform-Guided Multi-Scale Attention, or FTG MSA. This module uses both the amplitude of the four channel feature brightness map and the amplitude of the low-light image in its spatial domain as guidance information.

Lalam: And to make that guidance concrete, they combine the FFT of Y with the FFT of Fn to derive a guiding factor called Yb which is reshaped after an inverse Fourier transform over Pn. That’s how they use frequency domain information directly in the network.

Tom: So, it's not just applying a filter; it’s dynamically drawing relevant features from the source data during enhancement using both FFT and cross-attention mechanisms to guide what the model focuses on.

The paper's summary: Jane: When we look at their summary, they highlight that conventional low-light techniques, like histogram equalization or Retinex methods, are too laborious and time consuming because they often neglect the processing of missing information in image black spots. This makes them fail to meet the demands for human visibility and can even compromise performance when you need good texture details.

Lu: The paper argues that their MSFT approach is proposed to highlight those texture details of the low-light background while ensuring precise illumination recovery, which they show works by using the frequency domain amplitude spectrum of a normal light image added to the low-light image before reconstructing them via inverse Fourier transform.

Meng: They show visualization results where the former successfully achieves a brightness enhancement effect similar to that under normal lighting conditions, which gives a solid, realistic basis for the idea that we can prevent overexposure in these tough scenarios.

Lalam: That means they are showing that by using the frequency domain amplitude spectrum as guidance, they can achieve a brightness enhancement effect that looks realistic without blowing out highlights or losing detail.

Tom: It seems like the main takeaway here is moving beyond simple brightness correction and actually using the frequency domain to guide the network so it learns how to recover texture information instead of just guessing a uniform light level.

The paper's improvements: Jane: They suggest several improvements in their framework, starting with Multi-Shape Synergistic Attention, or MSSA. This module is designed to accommodate multi-scale feature extraction capabilities by having three serially connected components: MSA, SCSA, and a dynamic filter network.

Lu: Specifically, the MSA module is built by parallel-connecting Dilated Square Attention and Dilated Rectangle Attention in parallel to get better scale coverage. They also mention the necessity of the FFT module because without it, performance generally drops across all datasets.

Meng: And when we look at how they optimized this for deployment, they found that removing components like CMUNeXt significantly reduced parameters down to one point one million parameters and computational volume to about seventeen point four hundred fifty-four GFLOPs while keeping strong performance on the SDSD-outdoor and SDSD-indoor datasets.

Tom: That's a huge win for deployment, because they managed to keep the accuracy high while making the model much lighter than what we usually see in these complex architectures. The ablation study confirmed that removing the FTG module caused a general decline in performance across all tested datasets, proving its necessity.

Lalam: So, to summarize their improvements, they showed that all major components are needed; specifically, removing the FFT module caused a sharp drop because it’s indispensable for capturing and reconstructing global frequency information and structural consistency of the image.

Conclusion: Jane: So wrapping up this paper on "Low-light Image Enhancement via Multi-scale Attention combined with Fourier Transform," the main implication is that we can use a frequency domain guided attention mechanism to restore degraded images while effectively enhancing background details and accurately restoring the original illumination.

Tom: It’s a reference point for constructing models that use this frequency domain guided attention mechanism to restore images, which is really interesting because it gives us a new direction for how we approach image restoration problems.

Lu: The method does aim to effectively enhance the background details of low-light images while accurately restoring the original illumination, but they also flagged a limitation: the model is limited in dealing with images polluted by strong light due to ineffective removal of light pollution noise from the exposed part.

Meng: Also, it requires a large amount of paired data for training to perform well, so if you are trying to use this for an unsupervised augmentation field without paired data right now, it’s not ideal.

Lalam: And future work they suggest is exploring Attention calculations using both the amplitude spectrum and phase spectrum to see if that leads to even superior performance in frequency domain learning.

Tom: So, that's the rundown on MSFT; it’s a powerful method for enhancing low-light images when you have good data, but we still need to tackle those noise issues and scale up the training data requirements.

Wenbin Dua, Jian Longa, *, Zhu Caob,c,*

Key Laboratory of Smart Manufacturing in Energy Chemical Process, Ministry of Education, East China University of Science and Technology · College of Electronics and Information Engineering, Tongji University · Shanghai Research Institute for Intelligent Autonomous Systems, Tongji University

cs.CV

Submitted: 2026-07-27

Updated: 2026-07-27

Comments: 27 pages, 8 figures

Journal ref: Expert Systems with Applications 321, 132146 (2026)

DOI: 10.1016/j.eswa.2026.132146

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 77/100

The gist: Low-light image enhancement (LLIE) aims to improve image quality and clarity in diverse and demanding low-illumination environments.

Key concepts

Retinex decomposition
This is a technique used to separate an image into two components: the illumination (how bright the scene is) and the reflectance (the actual colors and details of objects). By separating these parts, the enhancement process can focus on correcting lighting issues independently from restoring image clarity.
Fourier Transform-Guided Multi-Scale Attention (FTG-MSA)
This core module uses frequency domain information to guide attention. It combines spatial brightness maps with low-light image amplitudes and uses FFT and cross-attention mechanisms to dynamically select and draw the most relevant features from the input data during enhancement.
Multi-Shape Synergistic Attention (MSSA)
This module is designed for multi-scale feature extraction, consisting of sequentially connected components like MSA, SCSA, and a dynamic filter network. It helps the model capture texture details across different scales while integrating information effectively into the feature space channels.

Terminology

Summary

Low-light image enhancement (LLIE) aims to improve image quality and clarity in diverse and demanding low-illumination environments. p. 0

Where the paper says it

Low-light Image Enhancement via Multi-scale Attention combined with Fourier Transform Wenbin Dua , Jian Longa,∗ , Zhu Caob,c,∗ aKey Laboratory of Smart Manufacturing in Energy Chemical Process, Ministry of Education, East China University of Science and Technology, Shanghai, 200237, China bCollege of Electronics and Information Engineering, Tongji University, Shanghai 201804, China cShanghai Research Institute for Intelligent Autonomous Systems, Tongji University, Shanghai 201210, China Abstract Low-light image enhancement (LLIE) aims to improve image quality and clarity in diverse and demanding low-illumination environments. However, existing deep learning-based LLIE methods struggle to accurately capture real-world illumination and restore texture details, largely because their algorithmic strengths remain underutilized.

This paper presents a supervised frequency domain deep learning network named multi-scale attention combined with the Fourier transform (MSFT) for LLIE. p. 0
Where the paper says it

However, existing deep learning-based LLIE methods struggle to accurately capture real-world illumination and restore texture details, largely because their algorithmic strengths remain underutilized. To address these issues, we present a supervised frequency domain deep learning network for LLIE, named multi-scale attention combined with the Fourier transform (MSFT) which adopts a U-shaped, one-stage architecture that infuses guidance from low-light images into the network by channeling it through multi-scale attention.

Methodology Overview

The proposed MSFT method is a two-stage U-shaped framework combining the Fourier transform with a CNN-Transformer Cross-mixed U-net framework. This architecture involves several key components including the Retinex decomposition performed in the original image space. p. 6

Where the paper says it

MSA is constructed by parallel-connecting DSA and DRA. Next, we systematically elaborate on the complete construction of the MSFT architecture, focusing on three key components: 1) The Retinex decomposition performed in the original image space.

The core enhancement module is the Fourier Transform-Guided Multi-Scale Attention (FTG-MSA). p. 2
Where the paper says it

Based on this, the enhancement weights can be adaptively adjusted according to the maximum gray value in the low-light area, thereby preventing the occurrence of overexposure and underexposure phenomena. Specifically, we design a multi-head self-attention module guided by the Fourier frequency-guided transform (FFG-MSA), which adds the amplitude of the four-channel feature brightness mape to the amplitude of the low-light image in the spatial domain as guidance information.

This module incorporates frequency domain information by combining the amplitude of the four-channel feature brightness map with the amplitude of the low-light image in the spatial domain as guidance information. p. 2
Where the paper says it

Based on this, the enhancement weights can be adaptively adjusted according to the maximum gray value in the low-light area, thereby preventing the occurrence of overexposure and underexposure phenomena. Specifically, we design a multi-head self-attention module guided by the Fourier frequency-guided transform (FFG-MSA), which adds the amplitude of the four-channel feature brightness mape to the amplitude of the low-light image in the spatial domain as guidance information.

The FTG-MSA module utilizes both FFT and cross-attention mechanisms to dynamically draw relevant features from the source data during enhancement. p. 3
Where the paper says it

the cross-attention mechanism can dynamically and selectively draw the most relevant features from the source data during enhancement, thereby enriching the image content. We conducted extensive experiments on five representative datasets: LOL, SID, SMID, and SDSD, and the subsets of the LOL and SDSD datasets were further divided.

Specifically, it combines FFT of Y with FFT of Fn to derive a guiding factor Yb = Reshape(IFFT(Ay + Afn, Pn)). p. 8
Where the paper says it

To effectively leverage the generated amplitude prior and to preserve the fundamental textural elements of the original image, we combine Ay and Afn to replace the amplitude. Next, we can derive the guiding factor and reshape it to Yb ∈ R HW×C through the application of the Inverse Fast Fourier Transform (IFFT), Yb = Reshape(IFFT(Ay + Afn , Pn)), (5) while the other branch is reshaped into tokens X ∈ R HW×C to calculate self-attention.

Synergistic Attention Components

The Multi-Shape Synergistic Attention (MSSA) module is designed to accommodate multi-scale feature extraction capabilities. MSSA consists of three serially connected modules: MSA [33], SCSA [32], and a dynamic filter network. p. 6

Where the paper says it

Figure 2(c) describes FTGT, comprising two Layer Normalization modules, one FTG-MSA module, and one Feed-Forward Network module. Figure 2(d) details MSSA, which integrates three serially connected modules (SCSA, MSA, and CMUNeXt).

The MSA module is constructed by parallel-connecting Dilated Square Attention (DSA) and Dilated Rectangle Attention (DRA) in parallel. p. 9
Where the paper says it

Subsequently, Fn3 proceeds through the SCSA module as F¯ n3 = ΦSCSA(Fn3 ), (10) where the SCSA module does not alter the channel dimensions of the training data. The MSA module consists of Dilated Square Attention (DSA) and Dilated Rectangle Attention (DRA) in parallel, aggregating information in square and rectangular regions respectively, to realize feature learning at different scales.

Performance Evaluation

Extensive experiments conducted on LOL, SID, SMID, and SDSD datasets demonstrate that MSFT significantly outperforms state-of-the-art competitors. For example, compared with Retinexformer, our method achieves a peak signal-to-noise ratio of up to 41. p. 0

Where the paper says it

Extensive experiments conducted on LOL, SID, SMID, and SDSD datasets demonstrate that MSFT significantly outperforms state-of-the-art competitors. For example, compared with Retinexformer, our method achieves a peak signal-to-noise ratio of up to 41.

76 decibels on the SDSD-outdoor dataset with an increase of 11. p. 0
Where the paper says it

For example, compared with Retinexformer, our method achieves a peak signal-to-noise ratio of up to 41. 76 decibels on the SDSD-outdoor dataset with an increase of 11.

92 decibels and a structural similarity index of 0. p. 0
Where the paper says it

76 decibels on the SDSD-outdoor dataset with an increase of 11. 92 decibels and a structural similarity index of 0.

988 with a 13. p. 0
Where the paper says it

92 decibels and a structural similarity index of 0. 988 with a 13.

80% improvement. p. 0
Where the paper says it

988 with a 13. 80% improvement.

The model achieved PSNRs of 41. p. 3
Where the paper says it

However, these methods largely overlook the influence of lighting conditions and lack effective strategies to address noise. In recent years, deep learning LLIE methods have commonly achieved better results than traditional hand-crafted methods.

76±0.03 and 39.04±0.02 on the SDSD-outdoor and SDSD-indoor datasets, respectively. p. 5
Where the paper says it

Comparison of Related Low-Light Enhancement Methods Method Algorithmic model Datasets Advantages Limitations RetinexNet [44] Decomposition → Adjustment → Reconstruction LOL and RAISE Integrates Retinex theory with neural networks Color distortion in color images Self-Calibrated Illumination (SCI) [45] Illumination Learning with Weight Sharing + SelfCalibrated Module + Noise Removal Modules + Unsupervised Training Loss MIT and LSRW Stable exposure, fast, flexible, and robust Low generalization ability URetinex-Net [17] Design learnable modules for initialization based on data dependencies, expansion optimization, and specification of lighting enhancement LOL and SICE Realize the noise suppression and detail preservation of the final decomposition result. Lack of effective processing for areas with uneven lighting EnlightenGAN [46] Global-Local Discriminators + U-Net Generator MEF, NPE, LIME, VV, DICM, BBD100k, and LOL Global-local synergy; balances authenticity and accurac Generalizability limited by training dataset PyDiff [47] Pyramid structure + Diffusion model LOL, LSRW , DICM, NPE, and VV Adaptability in multiple scenarios, rich content restoration High consumption of computing resources LightenDiff [48] Retinex decomposition of the latent space + illumination restoration of the unsupervised diffusion model LOLv2-real, LOLv2- synthetic Prevent the occurrence of global degradation and improve computational efficiency Poor adaptability of dynamic scenes Anlightendiff [49] A dynamical regulated diffusion anchoring mechanism LOL, VE-LOL, LOLv2-real, NPE, LIME, and VV Dynamically adjust the diffusion process to adapt to complex scenarios and suppress various noises The distortion of color content and the insufficient ability to restore details in high-contrast scenes SNR-Net [25] Transformer + CNN LOL series, SID, SMID, and SDSD (indoor/outdoor) Spatially adaptive enhancement Insufficient recovery of near-black regions IA-LLIE [21] Transformer + CNNs + U-net architecture LOLv1, LOLv2-real, LOLv2-synthetic, LSRW Superior lighting adaptation, true-to-life color recovery Insufficient noise robustness Retinexformer [51] One-stage Retinex framework + illumination-guided CNNsTransformer hybrid architecture LOL series, SID, SMID, and SDSD (indoor/outdoor) Avoids overexposure; preserves details and colors Reversal distortion in bright areas UHDformer [58] Cascade network + Fourier Transform LOLv1, LOLv2, and UHD-LL Excellent cross-scene generalization, superior denoising capability, and refined enhancement tailored for high-resolution images Not suitable for dynamic scenes SG-LLIE [53] Structure-Guided Transformer block + Hybrid StructureGuided Feature Extractor + U-net architecture LOL-v1, LOLv2-real, and NTIRE 2025 LLIE Robustness of image texture structure Poor noise robustness 6

Ablation Study Findings

Ablation experiments confirmed the necessity of all major components in the MSFT architecture. Removing FTGT led to a general decline in performance across all datasets. p. 8

Where the paper says it

Retinexformer [51] utilizes Transformer layers across multiple scales globally and fully exploits the potential of the attention mechanism by incorporating critical information. Refining the Retinexformer [51], we design a Fourier Transform-Guided Transformer (FTGT) to serve as a crucial restoration module.

The FFT module was found to be indispensable for capturing and reconstructing global frequency information and structural consistency of the image. p. 2
Where the paper says it

This approach leverages the observation that the frequency domain amplitude encodes the main brightness information, while the phase components relate to noise and structural details, as shown in Figure 1. Figure 1(a) shows the amplitude spectrum of a normal light image decomposed by FFT, which is added pixel-by-pixel to the amplitude spectrum of a low-light image.

Conclusion

We propose a Transformer-based method combined with the Fourier transform for low-light image enhancement. Our approach is capable of effectively enhancing the background details of low-light images while accurately restoring the original illumination. p. 2

Where the paper says it

To highlight the texture details of the low-light background, while ensuring precise illumination recovery, we propose a Multi-Scale attention combined with the Fourier Transform (MSFT) method. This approach leverages the observation that the frequency domain amplitude encodes the main brightness information, while the phase components relate to noise and structural details, as shown in Figure 1.

This framework provides a reference for constructing models that use a frequency-domain guided attention mechanism to restore degraded images. p. 5
Where the paper says it

Comparison of Related Low-Light Enhancement Methods Method Algorithmic model Datasets Advantages Limitations RetinexNet [44] Decomposition → Adjustment → Reconstruction LOL and RAISE Integrates Retinex theory with neural networks Color distortion in color images Self-Calibrated Illumination (SCI) [45] Illumination Learning with Weight Sharing + SelfCalibrated Module + Noise Removal Modules + Unsupervised Training Loss MIT and LSRW Stable exposure, fast, flexible, and robust Low generalization ability URetinex-Net [17] Design learnable modules for initialization based on data dependencies, expansion optimization, and specification of lighting enhancement LOL and SICE Realize the noise suppression and detail preservation of the final decomposition result. Lack of effective processing for areas with uneven lighting EnlightenGAN [46] Global-Local Discriminators + U-Net Generator MEF, NPE, LIME, VV, DICM, BBD100k, and LOL Global-local synergy; balances authenticity and accurac Generalizability limited by training dataset PyDiff [47] Pyramid structure + Diffusion model LOL, LSRW , DICM, NPE, and VV Adaptability in multiple scenarios, rich content restoration High consumption of computing resources LightenDiff [48] Retinex decomposition of the latent space + illumination restoration of the unsupervised diffusion model LOLv2-real, LOLv2- synthetic Prevent the occurrence of global degradation and improve computational efficiency Poor adaptability of dynamic scenes Anlightendiff [49] A dynamical regulated diffusion anchoring mechanism LOL, VE-LOL, LOLv2-real, NPE, LIME, and VV Dynamically adjust the diffusion process to adapt to complex scenarios and suppress various noises The distortion of color content and the insufficient ability to restore details in high-contrast scenes SNR-Net [25] Transformer + CNN LOL series, SID, SMID, and SDSD (indoor/outdoor) Spatially adaptive enhancement Insufficient recovery of near-black regions IA-LLIE [21] Transformer + CNNs + U-net architecture LOLv1, LOLv2-real, LOLv2-synthetic, LSRW Superior lighting adaptation, true-to-life color recovery Insufficient noise robustness Retinexformer [51] One-stage Retinex framework + illumination-guided CNNsTransformer hybrid architecture LOL series, SID, SMID, and SDSD (indoor/outdoor) Avoids overexposure; preserves details and colors Reversal distortion in bright areas UHDformer [58] Cascade network + Fourier Transform LOLv1, LOLv2, and UHD-LL Excellent cross-scene generalization, superior denoising capability, and refined enhancement tailored for high-resolution images Not suitable for dynamic scenes SG-LLIE [53] Structure-Guided Transformer block + Hybrid StructureGuided Feature Extractor + U-net architecture LOL-v1, LOLv2-real, and NTIRE 2025 LLIE Robustness of image texture structure Poor noise robustness 6

Although the model is limited in dealing with images polluted by strong light due to the ineffective removal of light pollution noise from the exposed part. p. 5
Where the paper says it

Comparison of Related Low-Light Enhancement Methods Method Algorithmic model Datasets Advantages Limitations RetinexNet [44] Decomposition → Adjustment → Reconstruction LOL and RAISE Integrates Retinex theory with neural networks Color distortion in color images Self-Calibrated Illumination (SCI) [45] Illumination Learning with Weight Sharing + SelfCalibrated Module + Noise Removal Modules + Unsupervised Training Loss MIT and LSRW Stable exposure, fast, flexible, and robust Low generalization ability URetinex-Net [17] Design learnable modules for initialization based on data dependencies, expansion optimization, and specification of lighting enhancement LOL and SICE Realize the noise suppression and detail preservation of the final decomposition result. Lack of effective processing for areas with uneven lighting EnlightenGAN [46] Global-Local Discriminators + U-Net Generator MEF, NPE, LIME, VV, DICM, BBD100k, and LOL Global-local synergy; balances authenticity and accurac Generalizability limited by training dataset PyDiff [47] Pyramid structure + Diffusion model LOL, LSRW , DICM, NPE, and VV Adaptability in multiple scenarios, rich content restoration High consumption of computing resources LightenDiff [48] Retinex decomposition of the latent space + illumination restoration of the unsupervised diffusion model LOLv2-real, LOLv2- synthetic Prevent the occurrence of global degradation and improve computational efficiency Poor adaptability of dynamic scenes Anlightendiff [49] A dynamical regulated diffusion anchoring mechanism LOL, VE-LOL, LOLv2-real, NPE, LIME, and VV Dynamically adjust the diffusion process to adapt to complex scenarios and suppress various noises The distortion of color content and the insufficient ability to restore details in high-contrast scenes SNR-Net [25] Transformer + CNN LOL series, SID, SMID, and SDSD (indoor/outdoor) Spatially adaptive enhancement Insufficient recovery of near-black regions IA-LLIE [21] Transformer + CNNs + U-net architecture LOLv1, LOLv2-real, LOLv2-synthetic, LSRW Superior lighting adaptation, true-to-life color recovery Insufficient noise robustness Retinexformer [51] One-stage Retinex framework + illumination-guided CNNsTransformer hybrid architecture LOL series, SID, SMID, and SDSD (indoor/outdoor) Avoids overexposure; preserves details and colors Reversal distortion in bright areas UHDformer [58] Cascade network + Fourier Transform LOLv1, LOLv2, and UHD-LL Excellent cross-scene generalization, superior denoising capability, and refined enhancement tailored for high-resolution images Not suitable for dynamic scenes SG-LLIE [53] Structure-Guided Transformer block + Hybrid StructureGuided Feature Extractor + U-net architecture LOL-v1, LOLv2-real, and NTIRE 2025 LLIE Robustness of image texture structure Poor noise robustness 6

It requires a large amount of paired data for training to perform well. p. 1
Where the paper says it

To address these issues, diffusion models exhibit exceptional capabilities in generating highly realistic and intricately detailed images. Most diffusion-based networks [26, 27] treat low-light images as image generation or image restoration tasks, which adopt supervised learning approaches with large-scale paired data to restore detailed image textures and optimize contrast.

Future Work

Future work suggests exploring Attention calculations using the amplitude spectrum and phase spectrum to further explore the superior performance of the Transformer in frequency-domain learning. Additionally, future work hopes to optimize the structure of diffusion models and reduce computational resources for its superior performance. p. 5

Where the paper says it

Comparison of Related Low-Light Enhancement Methods Method Algorithmic model Datasets Advantages Limitations RetinexNet [44] Decomposition → Adjustment → Reconstruction LOL and RAISE Integrates Retinex theory with neural networks Color distortion in color images Self-Calibrated Illumination (SCI) [45] Illumination Learning with Weight Sharing + SelfCalibrated Module + Noise Removal Modules + Unsupervised Training Loss MIT and LSRW Stable exposure, fast, flexible, and robust Low generalization ability URetinex-Net [17] Design learnable modules for initialization based on data dependencies, expansion optimization, and specification of lighting enhancement LOL and SICE Realize the noise suppression and detail preservation of the final decomposition result. Lack of effective processing for areas with uneven lighting EnlightenGAN [46] Global-Local Discriminators + U-Net Generator MEF, NPE, LIME, VV, DICM, BBD100k, and LOL Global-local synergy; balances authenticity and accurac Generalizability limited by training dataset PyDiff [47] Pyramid structure + Diffusion model LOL, LSRW , DICM, NPE, and VV Adaptability in multiple scenarios, rich content restoration High consumption of computing resources LightenDiff [48] Retinex decomposition of the latent space + illumination restoration of the unsupervised diffusion model LOLv2-real, LOLv2- synthetic Prevent the occurrence of global degradation and improve computational efficiency Poor adaptability of dynamic scenes Anlightendiff [49] A dynamical regulated diffusion anchoring mechanism LOL, VE-LOL, LOLv2-real, NPE, LIME, and VV Dynamically adjust the diffusion process to adapt to complex scenarios and suppress various noises The distortion of color content and the insufficient ability to restore details in high-contrast scenes SNR-Net [25] Transformer + CNN LOL series, SID, SMID, and SDSD (indoor/outdoor) Spatially adaptive enhancement Insufficient recovery of near-black regions IA-LLIE [21] Transformer + CNNs + U-net architecture LOLv1, LOLv2-real, LOLv2-synthetic, LSRW Superior lighting adaptation, true-to-life color recovery Insufficient noise robustness Retinexformer [51] One-stage Retinex framework + illumination-guided CNNsTransformer hybrid architecture LOL series, SID, SMID, and SDSD (indoor/outdoor) Avoids overexposure; preserves details and colors Reversal distortion in bright areas UHDformer [58] Cascade network + Fourier Transform LOLv1, LOLv2, and UHD-LL Excellent cross-scene generalization, superior denoising capability, and refined enhancement tailored for high-resolution images Not suitable for dynamic scenes SG-LLIE [53] Structure-Guided Transformer block + Hybrid StructureGuided Feature Extractor + U-net architecture LOL-v1, LOLv2-real, and NTIRE 2025 LLIE Robustness of image texture structure Poor noise robustness 6

Future research will also explore how to apply attention in the frequency domain. p. 0
Where the paper says it

However, existing deep learning-based LLIE methods struggle to accurately capture real-world illumination and restore texture details, largely because their algorithmic strengths remain underutilized. To address these issues, we present a supervised frequency domain deep learning network for LLIE, named multi-scale attention combined with the Fourier transform (MSFT) which adopts a U-shaped, one-stage architecture that infuses guidance from low-light images into the network by channeling it through multi-scale attention.

Acknowledgements

This work was partially supported by various grants including Grant 25ZR1402098 and Grant YH0142234.

References

The paper references numerous related works in low-light image enhancement methods. These include traditional techniques like gamma correction and histogram equalization. p. 3

Where the paper says it

LLIE aims to improve the quality of low-visibility images that suffer from poor lighting conditions. Traditional methods for low-light enhancement mainly rely on gamma correction [35, 36], histogram equalization [37, 38, 39], and traditional Retinex theory [40, 41] to enhance low-light images.

Quantitative Results

MSFT significantly outperforms existing state-of-the-art methods across various metrics. In contrast to synthetic images, its enhancement of extremely dark background also produces a few artifacts, as shown in Figure 7. p. 5

Where the paper says it

Comparison of Related Low-Light Enhancement Methods Method Algorithmic model Datasets Advantages Limitations RetinexNet [44] Decomposition → Adjustment → Reconstruction LOL and RAISE Integrates Retinex theory with neural networks Color distortion in color images Self-Calibrated Illumination (SCI) [45] Illumination Learning with Weight Sharing + SelfCalibrated Module + Noise Removal Modules + Unsupervised Training Loss MIT and LSRW Stable exposure, fast, flexible, and robust Low generalization ability URetinex-Net [17] Design learnable modules for initialization based on data dependencies, expansion optimization, and specification of lighting enhancement LOL and SICE Realize the noise suppression and detail preservation of the final decomposition result. Lack of effective processing for areas with uneven lighting EnlightenGAN [46] Global-Local Discriminators + U-Net Generator MEF, NPE, LIME, VV, DICM, BBD100k, and LOL Global-local synergy; balances authenticity and accurac Generalizability limited by training dataset PyDiff [47] Pyramid structure + Diffusion model LOL, LSRW , DICM, NPE, and VV Adaptability in multiple scenarios, rich content restoration High consumption of computing resources LightenDiff [48] Retinex decomposition of the latent space + illumination restoration of the unsupervised diffusion model LOLv2-real, LOLv2- synthetic Prevent the occurrence of global degradation and improve computational efficiency Poor adaptability of dynamic scenes Anlightendiff [49] A dynamical regulated diffusion anchoring mechanism LOL, VE-LOL, LOLv2-real, NPE, LIME, and VV Dynamically adjust the diffusion process to adapt to complex scenarios and suppress various noises The distortion of color content and the insufficient ability to restore details in high-contrast scenes SNR-Net [25] Transformer + CNN LOL series, SID, SMID, and SDSD (indoor/outdoor) Spatially adaptive enhancement Insufficient recovery of near-black regions IA-LLIE [21] Transformer + CNNs + U-net architecture LOLv1, LOLv2-real, LOLv2-synthetic, LSRW Superior lighting adaptation, true-to-life color recovery Insufficient noise robustness Retinexformer [51] One-stage Retinex framework + illumination-guided CNNsTransformer hybrid architecture LOL series, SID, SMID, and SDSD (indoor/outdoor) Avoids overexposure; preserves details and colors Reversal distortion in bright areas UHDformer [58] Cascade network + Fourier Transform LOLv1, LOLv2, and UHD-LL Excellent cross-scene generalization, superior denoising capability, and refined enhancement tailored for high-resolution images Not suitable for dynamic scenes SG-LLIE [53] Structure-Guided Transformer block + Hybrid StructureGuided Feature Extractor + U-net architecture LOL-v1, LOLv2-real, and NTIRE 2025 LLIE Robustness of image texture structure Poor noise robustness 6

The parameter count is relatively low at 1. p. 3
Where the paper says it

The deep learning approach designs networks based on specific learnable parameter matrices and optical theories such as Retinex. This enables the network to learn ideal image enhancement priors, thereby accurately achieving low-light enhancement, exposure stabilization, and other functions.

25 million parameters. p. 8
Where the paper says it

When Fl is fed as input to FTGT, the guidance information Fn also follows subsequently. Fn utilized undersampled convolution with the same parameters as Fl to obtain the corresponding rank features Fni , where i = 0, 1, 2.

The absence of the FFT module caused a sharp drop in performance on some datasets. p. 5
Where the paper says it

Comparison of Related Low-Light Enhancement Methods Method Algorithmic model Datasets Advantages Limitations RetinexNet [44] Decomposition → Adjustment → Reconstruction LOL and RAISE Integrates Retinex theory with neural networks Color distortion in color images Self-Calibrated Illumination (SCI) [45] Illumination Learning with Weight Sharing + SelfCalibrated Module + Noise Removal Modules + Unsupervised Training Loss MIT and LSRW Stable exposure, fast, flexible, and robust Low generalization ability URetinex-Net [17] Design learnable modules for initialization based on data dependencies, expansion optimization, and specification of lighting enhancement LOL and SICE Realize the noise suppression and detail preservation of the final decomposition result. Lack of effective processing for areas with uneven lighting EnlightenGAN [46] Global-Local Discriminators + U-Net Generator MEF, NPE, LIME, VV, DICM, BBD100k, and LOL Global-local synergy; balances authenticity and accurac Generalizability limited by training dataset PyDiff [47] Pyramid structure + Diffusion model LOL, LSRW , DICM, NPE, and VV Adaptability in multiple scenarios, rich content restoration High consumption of computing resources LightenDiff [48] Retinex decomposition of the latent space + illumination restoration of the unsupervised diffusion model LOLv2-real, LOLv2- synthetic Prevent the occurrence of global degradation and improve computational efficiency Poor adaptability of dynamic scenes Anlightendiff [49] A dynamical regulated diffusion anchoring mechanism LOL, VE-LOL, LOLv2-real, NPE, LIME, and VV Dynamically adjust the diffusion process to adapt to complex scenarios and suppress various noises The distortion of color content and the insufficient ability to restore details in high-contrast scenes SNR-Net [25] Transformer + CNN LOL series, SID, SMID, and SDSD (indoor/outdoor) Spatially adaptive enhancement Insufficient recovery of near-black regions IA-LLIE [21] Transformer + CNNs + U-net architecture LOLv1, LOLv2-real, LOLv2-synthetic, LSRW Superior lighting adaptation, true-to-life color recovery Insufficient noise robustness Retinexformer [51] One-stage Retinex framework + illumination-guided CNNsTransformer hybrid architecture LOL series, SID, SMID, and SDSD (indoor/outdoor) Avoids overexposure; preserves details and colors Reversal distortion in bright areas UHDformer [58] Cascade network + Fourier Transform LOLv1, LOLv2, and UHD-LL Excellent cross-scene generalization, superior denoising capability, and refined enhancement tailored for high-resolution images Not suitable for dynamic scenes SG-LLIE [53] Structure-Guided Transformer block + Hybrid StructureGuided Feature Extractor + U-net architecture LOL-v1, LOLv2-real, and NTIRE 2025 LLIE Robustness of image texture structure Poor noise robustness 6

The FFT module is indispensable for capturing and reconstructing global frequency information and structural consistency of the image. p. 2
Where the paper says it

This approach leverages the observation that the frequency domain amplitude encodes the main brightness information, while the phase components relate to noise and structural details, as shown in Figure 1. Figure 1(a) shows the amplitude spectrum of a normal light image decomposed by FFT, which is added pixel-by-pixel to the amplitude spectrum of a low-light image.

Limitations

The method is limited in dealing with images polluted by strong light due to the ineffective removal of light pollution noise from the exposed part. It also requires a large amount of paired data for training to perform well. p. 1

Where the paper says it

To address these issues, diffusion models exhibit exceptional capabilities in generating highly realistic and intricately detailed images. Most diffusion-based networks [26, 27] treat low-light images as image generation or image restoration tasks, which adopt supervised learning approaches with large-scale paired data to restore detailed image textures and optimize contrast.

The model is not suitable for an unsupervised augmentation field without paired data at present. p. 6
Where the paper says it

Comparison of Related Low-Light Enhancement Methods Method Algorithmic model Datasets Advantages Limitations FourierDiff [60] Fourier priors-guided diffusion model + Spatial-frequency alternating optimization LOL-blur, RealBlur Clear, consistent, and independence of paired images. Fails without content priors and under extremely low illumination DMFourLLIE [61] Fourier reconstruction + Fourier convolution + multiscales convolution LSRW-Huawei, LOLv2-real, LOL-v2- synthesis, LIME, VV, DICM, NPE, and MEF High-fidelity finegrained texture and spatial structure Only suitable for reconstruction in static scenes 3.

It also does not take into account images acquired in scenarios with severe noise pollution. p. 1
Where the paper says it

These models mainly use deep neural networks [20] frameworks to learn to manipulate brightness, color, tone, and contrast information from normal light maps to improve image quality. However, mainstream deep learning methods [21, 22, 23] are limited in their ability to adaptively adjust the brightness distribution based on the texture information present in an image, as they fail to take into account the regional distribution of the texture information.

Future work will explore applying attention calculations using the amplitude spectrum and phase spectrum to further explore the superior performance of the Transformer in frequency-domain learning. Future work will also explore how to apply attention in the frequency domain. p. 4

Where the paper says it

As shown in Table 1 of the above-mentioned related work. Inspired by the above work in the frequency domain, we extract multi-scale frequency features to insert into the corresponding scale self-attention.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Multi-scale guidance through Fourier Transform can be used to prevent overexposure and enhance dim backgrounds simultaneously by channeling information through multi-scale attention guided by frequency domain amplitude and phase components.

  2. The model can effectively capture texture details of the low-light background while preserving realistic illumination, as demonstrated by the MSFT method achieving a peak signal-to-noise ratio of up to 41.76 decibels on the SDSD-outdoor dataset with an increase of 11.92 decibels and a structural similarity index of 0.988 with a 13.80% improvement compared to Retinexformer.

  3. The system can be adapted to handle uneven illumination in low-light images by designing a lightweight network that effectively integrates information in high-dimensional space into feature space channels containing rich texture information, as described by the multi-shape synergistic attention module.

  4. The MSFT framework can achieve high fidelity and richly detailed results through its two-stage U-shaped architecture combining CNN efficiency with Transformer's global context modeling, ensuring a balance between precise understanding of global context and perfect reconstruction of local details.

  5. The system can dynamically and selectively draw the most relevant features from source data during enhancement using the FTG-MSA mechanism, which incorporates frequency domain information to enrich image content during restoration.

  6. The proposed architecture can be optimized for resource efficiency by utilizing a lightweight design within MSSA, where removing components like CMUNeXt significantly reduced parameters to 1.101MParams and computational volume to 17.454GFLOPs while maintaining strong performance on SDSD-outdoor and SDSD-indoor datasets.

  7. The system can be improved by exploring the superior performance of the Transformer in frequency domain learning by performing Attention calculations using the amplitude spectrum and phase spectrum, potentially leading to better restoration of degraded images.

Abstract

Low-light image enhancement (LLIE) aims to improve image quality and clarity in diverse and demanding low-illumination environments. However, existing deep learning-based LLIE methods struggle to accurately capture real-world illumination and restore texture details, largely because their algorithmic strengths remain underutilized. To address these issues, we present a supervised frequency domain deep learning network for LLIE, named multi-scale attention combined with the Fourier transform (MSFT) which adopts a U-shaped, one-stage architecture that infuses guidance from low-light images into the network by channeling it through multi-scale attention. We further fuse the amplitude information from priori channels with that of the low-light image in MSFT's self-created module, and carry out multi-scale guidance along with the network. Subsequently, to better enhance the faint feature, such as fine content and textures, and to better fuse global context confidence in the decoding stage, we separately introduce a multi-shape synergistic attention and a lightweight network that effectively integrate information in high-dimensional space to embed into the superlative feature space channel containing rich texture information. Extensive experiments conducted on LOL, SID, SMID, and SDSD datasets demonstrate that MSFT significantly outperforms state-of-the-art competitors. For example, compared with Retinexformer, our method achieves a peak signal-to-noise ratio of up to 41.76 decibels on the SDSD-outdoor dataset with an increase of 11.92 decibels and a structural similarity index of 0.988 with a 13.80% improvement.

Sources

Related papers