RetinexDual: Retinex-based Dual Nature Approach for Generalized Ultra-High-Definition Image Restoration

arXiv:2508.04797 · cs.CV · Submitted 2025-08-06 · 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: "RetinexDual: Retinex-based Dual Nature Approach for Generalized Ultra-High-Definition Image Restoration".

Jane: Here's the gist of the paper. RetinexDual presents a novel Retinex theory-based framework designed for generalized Ultra-High-Definition Image Restoration by leveraging two complementary sub-networks: Scale Attentive maMBA (SAMBA) and the Frequency Illumination Adaptor (FIA).

Tom: Problem Motivation Traditional methods for Ultra-High-Definition Image Restoration encounter significant drawbacks, as extreme downsampling induces irreversible information loss in UHD images, while pure frequency-domain approaches are ineffective for spatially confined image artifacts because of the loss of degradation locality.

Jane: Research has shown that for global haze, the frequency domain exhibits a higher PSNR compared to the spatial domain between the ground truth and input image, suggesting that the frequency domain more effectively captures the uniform degradation pattern.

Tom: Conversely, for localized blur, the spatial domain shows a higher PSNR compared to the frequency domain, indicating that the spatial domain better reflects the localized degradation characteristics.

Paper summary: Tom: So, to wrap up what we just heard about "RetinexDual: Retinex-based Dual Nature Approach for Generalized Ultra-High-Definition Image Restoration," the authors are proposing a framework based on retinex theory to solve generalized ultra-high-definition image restoration problems.

Jane: They claim this approach overcomes the limitations of previous methods by using two complementary sub-networks, SAMBA and FIA, to tackle the reflectance and illumination components separately.

Tom: SAMBA is responsible for correcting distortions in the reflectance component, employing a scale-adaptive mamba block to handle artifacts caused by degradation in that part of the image.

Lu: And FIA handles the illumination aspect by operating in the frequency domain, using a Fourier correction block that decomposes features into amplitude and phase before reassembling them with an inverse fast fourier transform.

Meng: It’s interesting how they designed SAMBA with a coarse-to-fine mechanism to overcome the causal modeling issues inherent in mamba, which is something that could be a pain when you're training complex models.

Lalam: And they use a joint objective function for training, which includes losses like the contrastive loss and the perceptual loss, plus some other things to make sure it generalizes well across different restoration tasks.

Tom: The paper also highlights their multi-level training strategy using three different scales of output images, I1, I2, and I3 to enhance the restoration quality at different resolutions.

Jane: It seems like they are showing that this separation of concerns between reflectance correction and illumination adjustment is key to achieving better results in these UHD restoration tasks.

Lu: If you look at their comparative results across different tasks, like dehazing and deblurring, they show state-of-the-art performance.

Meng: But we have to remember that this is a theoretical framework for now; it’s not a finished product ready to deploy on massive scale yet.

Conclusion: Tom: So, looking at the whole picture of "RetinexDual: Retinex-based Dual Nature Approach for Generalized Ultra-High-Definition Image Restoration," the main thing is that they’ve successfully used retinex theory as a general concept to build a dual architecture.

Jane: They’ve shown that addressing the duality of reflectance and illumination separately through SAMBA and FIA leads to state-of-the-art performance in several UHD restoration tasks.

Lu: The authors' main contribution is proposing this novel RetinexDual framework, which handles the complexity of ultra-high-definition images by splitting the problem into two distinct networks.

Meng: It’s an interesting direction for representation learning, but we still need to see how this translates into something that can run efficiently in a production environment.

Lalam: The paper confirms that RetinexDual achieves state-of-the-art results on four different UHD IR tasks based on their comparative evaluations.

Tom: So, while the performance numbers are impressive, the authors themselves acknowledge that they haven't optimized this for inference speed and size yet.

Jane: They conclude by suggesting that future work should focus on optimizing inference time and tackling challenges in restoring certain difficult images.

Department of Electrical & Computer Engineering, McMaster University

cs.CV

Submitted: 2025-08-06

Updated: 2025-11-19

DOI: 10.1007/978-3-032-31335-5_7

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

Importance score: 90/100

The gist: RetinexDual presents a novel Retinex theory-based framework designed for generalized Ultra-High-Definition Image Restoration by leveraging two complementary sub-networks…

Key concepts

Retinex Theory
This theory breaks an image into two parts: reflectance (R), which represents the intrinsic properties of the scene, and illumination (L), which represents how light affects that scene. The goal is to separate these components to restore a degraded image by correcting both the scene's details and the lighting conditions independently.
Scale Attentive maMBA (SAMBA)
This network focuses on fixing artifacts in the reflectance component of an image. It uses a specialized Mamba block that adapts its processing based on different scales of features, allowing it to effectively correct complex distortions and restore fine details in the image structure.
Frequency Illumination Adaptor (FIA)
This part handles the illumination component by working in the frequency domain. It uses a Fourier Correction Block to analyze and adjust color and lighting distortions globally, ensuring precise correction of how light has affected the image's overall appearance.

Terminology

Summary

RetinexDual presents a novel Retinex theory-based framework designed for generalized Ultra-High-Definition Image Restoration by leveraging two complementary sub-networks: Scale Attentive maMBA (SAMBA) and the Frequency Illumination Adaptor (FIA). p. 0

Where the paper says it

To overcome these limitations, we present RetinexDual, a novel Retinex theory-based framework designed for generalized UHD IR tasks. RetinexDual leverages two complementary sub-networks: the Scale-Attentive maMBA (SAMBA) and the Frequency Illumination Adaptor (FIA).

Problem Motivation

Traditional methods for Ultra-High-Definition Image Restoration encounter significant drawbacks, as extreme downsampling induces irreversible information loss in UHD images, while pure frequency-domain approaches are ineffective for spatially confined image artifacts because of the loss of degradation locality. Research has shown that for global haze, the frequency domain exhibits a higher PSNR compared to the spatial domain between the ground truth and input image, suggesting that the frequency domain more effectively captures the uniform degradation pattern. p. 1

Where the paper says it

We observed the difference between the clear image and the degraded image in both spatial and frequency domains, as shown in Fig. For global haze, the frequency domain exhibits a higher PSNR compared to the spatial domain between the ground truth and input image, suggesting that the frequency domain more effectively captures the uniform degradation pattern.

Conversely, for localized blur, the spatial domain shows a higher PSNR compared to the frequency domain, indicating that the spatial domain better reflects the localized degradation characteristics. p. 1
Where the paper says it

For global haze, the frequency domain exhibits a higher PSNR compared to the spatial domain between the ground truth and input image, suggesting that the frequency domain more effectively captures the uniform degradation pattern. Conversely, for localized blur, the spatial domain shows a higher PSNR compared to the frequency domain, indicating that the spatial domain better reflects the localized degradation characteristics.

Framework Architecture

RetinexDual decomposes an image into a reflection component R and an illumination component L such that I = R ⊙ L. The framework utilizes two main networks:

  1. Scale Attentive maMBA (SAMBA), which focuses on correcting the artifacts and distortions caused by degradation in the reflectance component by employing a tailored Scale-Adaptive Mamba block. p. 2

    Where the paper says it

    RetinexDual consists of two main networks: Scale Attentive maMBA (SAMBA) and Frequency Illumination Adaptor (FIA). SAMBA focuses on correcting the artifacts and distortions caused by degradation in the reflectance component by employing our tailored Scale-Adaptive Mamba block.

  2. Frequency Illumination Adaptor (FIA), which ensures precise correction of color and illumination distortions by operating in the frequency domain and leveraging the global context provided by it. p. 0

    Where the paper says it

    SAMBA, responsible for correcting the reflectance component, utilizes a coarse-to-fine mechanism to overcome the causal modeling of mamba, which effectively reduces artifacts and restores intricate details. On the other hand, FIA ensures precise correction of color and illumination distortions by operating in the frequency domain and leveraging the global context provided by it.

Component Details

The SAMBA branch is designed with a coarse-to-fine mechanism to overcome the causal modeling of mamba, which effectively reduces artifacts and restores intricate details. Specifically, the Scale Adaptive Mamba Block (SAMB) starts with a Residual Diluted Block (RDB) for feature refining and extraction, followed by three multi-scale levels of the feature map processed by diluted convolution and passed to the Group State Space Block (GSSB). p. 5

Where the paper says it

Our SAMB starts with a Residual Diluted Block (RDB) for feature refining and extraction. Then, our second contribution lies in using three multi-scale levels of the feature map (1, 1/2, 1/4), which are further processed by diluted convolution, concatenated together, and passed to the Group State Space Block (GSSB) eventually.

The Group State Space Module (GSSM) introduces local El and global Eg embeddings to represent pixels with similar semantics into the C component of the state space model, overcoming the causal nature of Mamba. p. 6
Where the paper says it

Given x ∈ R H×W×(3C) , we start by using positional encoding [7] on x to preserve the spatial information structure. Aiming to overcome the causal nature of Mamba, we introduce local El and global Eg embeddings in the block and across the model, respectively, to represent pixels with similar semantics into C component of the state space model as image-specific embedding E = (ElEg)Ycp.

The FIA branch operates in the Fourier domain to alter the illumination component, featuring a Fourier Correction Block (FCB) which uses FFT to decompose input features into amplitude A and phase P, processes them with two 1×1 convolutions, and reassembles the image using IFFT. p. 4

Where the paper says it

RetinexDual 5 Scale Attentive MaMBA (SAMBA) Frequency Illumination Adaptor (FIA) FCB FCB FCB FCB FCB FCB SAMB SAMB SAMB SAMB SAMB SAMB + + + + + + p p Input Multi-level Outputs 2 × Down 4 × Down 2 × Down 4 × Down SAMB FCB + p Scale-Adaptive Mamba Block Fourier Correction Block Conv 1×1 Depth-wise Conv 5×5 Addition Piling Tensors Element Multiplication LN FFT Amplitude Phase Relu Relu IFFT + Scale Fourier Correction Block (FCB) Conv 3×3 LN Layer Normalization Residual Diluted Block (1,2,1) 3×3 Feature Map ³×2 ³×4 Diluted Block (1,2,3,2,1) 3×3 C ±×2 ±×4 GSSB Scale Adaptive Mamba Block (SAMB) × N GSSB Grouping State Space Block S + S Split Fig. 2: RetinexDual Overview.

This compact design includes only 0. p. 7
Where the paper says it

To avoid instability and apply adaptive scaling, we multiply xˆ with the input and use a final 3 × 3 convolution layer followed by a scaled skip connection. Maintaining such a compact design that includes only 0.

2M parameters. p. 7
Where the paper says it

Maintaining such a compact design that includes only 0. 2M parameters serves as our third contribution.

Training Strategy

The model employs a multi-level training objective utilizing four different loss functions to enhance generalization across various UHD IR tasks. The objective is defined as L = λcbLcb(I, ˆI) + λF F TL1(F F T(I), F F T(ˆI)) + λssimLssim(I, ˆI) + λpLp(I, ˆI). p. 7

Where the paper says it

4 Multi-level Training Objective: Aiming for generalization over all UHD IR tasks, We utilize four different training objectives to enhance our network’s performance: Charbonnier loss Lcb [13], FFT loss LF F T , SSIM loss Lssim [28], and perceptual loss Lp [39] using VGG16 network [20]. Given I, ˆI refer to the ground truth and predicted image, respectively, the training objective is defined as follows: L = λcbLcb(I, ˆI) + λF F TL1(F F T(I), F F T( ˆI)) +λssimLssim(I, ˆI) + λpLp(I, ˆI), (8) where λcb, λF F T , λssim, λp are weight factors for each of the losses with 1, 0.

Furthermore, a multi-level loss Lmulti-level is applied by having output from multiple levels of the framework with different scales I1, I2, I3 as shown in Fig. p. 7
Where the paper says it

4, respectively. Aiming to inject the optimization into the low levels of the model, we follow the deep supervision strategy by having output from multiple levels of the framework with different scales I1, I2, I3 as shown in Fig.

2.

Performance Evaluation

RetinexDual outperforms existing methods in four UHD IR tasks qualitatively and quantitatively. In the Low-Light Image Enhancement results, RetinexDual achieves more effectiveness for the task, outperforming previous SOTA methods by 4. p. 9

Where the paper says it

The remaining methods, such as UHDFour [15], Wave-Mamba [43], UHDFormer [24], ERR [40], and RetinexDual demonstrate more effectiveness for the task. While our model outperforms the previous SOTA method by 4.

7% in PSNR. p. 1
Where the paper says it

Kishawy et al. DehazeInput GT PSNR: 14.

For Image Deraining, RetinexDual still achieved the highest score among all models. p. 9
Where the paper says it

Quantitative results in Table 2 show slight differences in terms of PSNR and SSIM between UDR-Mixer, ERR, and our model. However, RetinexDual still achieved the highest score among all models, in addition to delivering the best visual quality, especially in a challenging scene as shown in Fig.

Similarly, in Image Dehazing Results, RetinexDual significantly outperforms SOTA methods by 6% in terms of PSNR. p. 9
Where the paper says it

Table 3 summarizes the results between our model and six other approaches, including UHD [42], Restormer [38], Uformer [27], DehazeFormer [21], UHDFormer [24], D2Net [30], and ERR [40]. While RetinexDual significantly outperforms SOTA methods by 6% in terms of PSNR, and 0.

The ablation studies confirm the importance of employing distinct designs for each branch and the effectiveness of its various components. p. 0
Where the paper says it

Evaluating RetinexDual on four UHD IR tasks, namely deraining, deblurring, dehazing, and Low-Light Image Enhancement (LLIE), shows that it outperforms recent methods qualitatively and quantitatively. Ablation studies demonstrate the importance of employing distinct designs for each branch in RetinexDual, as well as the effectiveness of its various components.

Conclusion

RetinexDual is a novel approach that utilizes retinex theory as a generalized IR concept and addresses the duality nature of its decomposition separately using two complementary sub-networks. It achieves state-of-the-art performance in four UHD IR tasks, although it is not optimized in terms of inference time and size. p. 1

Where the paper says it

In another aspect, Retinexformer [2] proposed a retinex theory-based model to restore low-light images, and multiple researchers have followed the same concept [1,16]. Although these methods demonstrated state-of-the-art performance,

Future work should focus on optimizing inference time and addressing challenges in restoring certain difficult images. p. 1
Where the paper says it

However, following such a "DownsamplingEnhancement-Upsampling" modeling paradigm, crucial details are removed due to the separation between resampling operations and the enhancement process [36]. Furthermore, Wave-Mamba [43], and ERR [40] focused on restoring UHD images by operating on various frequency bands to leverage the global context and the simplified analysis of certain frequency-based patterns achieved in the frequency domain.

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. Improved AI system can perform unified UHD Image Restoration by decoupling image decomposition into reflectance and illumination components, as RetinexDual decomposes the image into reflection component R and illumination component L.

  2. The system utilizes Scale Attentive maMBA (SAMBA) to correct reflectance artifacts using a coarse-to-fine mechanism, which is described as overcoming the causal modeling of mamba, which effectively reduces artifacts and restores intricate details.

  3. The system employs the Frequency Illumination Adaptor (FIA) to handle color and illumination distortions by operating in the Fourier domain, leveraging the global context provided by it for precise correction.

  4. SAMBA incorporates a Scale Adaptive Mamba Block (SAMB) that adapts its scan strategy based on feature relations across different scales, allowing it to capture the relationship between features far from each other in the image, overcoming the drawback of convolution.

  5. FIA includes a Fourier Correction Block (FCB) with a reduced architectural complexity without compromising performance, which processes input features via FFT to decompose them into amplitude and phase components.

  6. The training objective is enhanced by a multi-level approach using four different losses including Charbonnier loss, FFT loss, SSIM loss, and perceptual loss to achieve generalization across various restoration tasks.

  7. The system can be trained using a deep supervision strategy where each scale is reconstructed following the retinex theory to avoid violating its concept by applying bilinear interpolation on ground truth at different scales.

Abstract

Advancements in image sensing have elevated the importance of Ultra-High-Definition Image Restoration (UHD IR). Traditional methods, such as extreme downsampling or transformation from the spatial to the frequency domain, encounter significant drawbacks: downsampling induces irreversible information loss in UHD images, while our frequency analysis reveals that pure frequency-domain approaches are ineffective for spatially confined image artifacts, primarily due to the loss of degradation locality. To overcome these limitations, we present RetinexDual, a novel Retinex theory-based framework designed for generalized UHD IR tasks. RetinexDual leverages two complementary sub-networks: the Scale-Attentive maMBA (SAMBA) and the Frequency Illumination Adaptor (FIA). SAMBA, responsible for correcting the reflectance component, utilizes a coarse-to-fine mechanism to overcome the causal modeling of mamba, which effectively reduces artifacts and restores intricate details. On the other hand, FIA ensures precise correction of color and illumination distortions by operating in the frequency domain and leveraging the global context provided by it. Evaluating RetinexDual on four UHD IR tasks, namely deraining, deblurring, dehazing, and Low-Light Image Enhancement (LLIE), shows that it outperforms recent methods qualitatively and quantitatively. Ablation studies demonstrate the importance of employing distinct designs for each branch in RetinexDual, as well as the effectiveness of its various components.

Sources

Related papers