Spatial-frequency-scale collaborative guided diffusion method for low-light image enhancement

Abstract

This invention belongs to the field of image processing technology and relates to a spatial-frequency-scale coordinated guided diffusion method for low-light image enhancement. First, structural edge extraction and wavelet decomposition are performed on the input low-light image to construct a two-stage wavelet structure prior guidance module. Then, a global scale position embedding is generated by jointly encoding the diffusion time step, spatial coordinates, and image scale information, and the structural conditions and position embeddings are jointly injected into a conditional diffusion network. During the back-diffusion process, a residual frequency-domain phase mixer is introduced to jointly model the amplitude spectrum and phase spectrum, achieving coordinated enhancement of spatial and frequency domain features. Finally, the enhanced normal-exposure image is reconstructed through multi-layer feature decoding. This invention effectively solves the problems of insufficient structure restoration, spatial discontinuity, and loss of texture details in existing low-light image enhancement methods, achieving coordinated optimization of noise suppression, structure preservation, and texture restoration, and has good engineering application value and promotion prospects.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to a spatial-frequency-scale coordinated guided diffusion method for low-light image enhancement. Background Technology

[0002] Low-light image enhancement aims to restore the brightness, color, and texture details of images in low-light environments, and is a fundamental technology in fields such as computer vision, intelligent security, autonomous driving, and intelligent terminal imaging. Existing methods mainly include traditional enhancement methods, deep learning methods, frequency domain enhancement methods, and diffusion model-based methods.

[0003] Traditional methods mainly include histogram equalization and Retinex theory, which enhance images by adjusting the brightness distribution or separating the illumination and reflection components, but they are prone to problems such as noise amplification, color distortion and loss of detail.

[0004] Deep learning methods utilize convolutional neural networks or Transformers to learn the mapping relationship between low-light images and normal-light images, which can achieve good enhancement effects. However, in extremely low-light scenes, there are still problems such as insufficient structure restoration, texture blurring, and limited generalization ability.

[0005] Frequency domain methods incorporate frequency information through wavelet transform and Fourier transform to achieve brightness restoration and detail enhancement. Wavelet transform effectively separates noise from structural information, while Fourier transform utilizes the amplitude and phase spectra for frequency domain modeling. However, most existing methods focus on amplitude spectrum modulation, underutilizing the phase spectrum, which contains rich structural information.

[0006] Diffusion models have made significant progress in image enhancement due to their powerful distribution modeling capabilities and progressive denoising mechanisms. However, existing diffusion models often lack explicit structural constraints and are limited by fixed-size image patch training methods, which can easily lead to spatial discontinuities and block artifacts during high-resolution image inference.

[0007] Overall, low-light image enhancement technology has gradually evolved from traditional image processing methods to a comprehensive enhancement framework that integrates deep learning, frequency domain modeling, and diffusion generation models. Different technical approaches have their own characteristics in terms of brightness restoration, detail reconstruction, structure preservation, and visual perception quality, laying the foundation for the development of low-light image enhancement technology.

[0008] While existing low-light image enhancement techniques have made some progress in brightness restoration and visual quality improvement, they still face challenges in extremely low-light scenes, including insufficient structure preservation, lack of spatial consistency, and inadequate utilization of frequency domain features. These challenges make it difficult to simultaneously achieve noise suppression, detail restoration, and global visual consistency. Specifically, these issues manifest in the following ways:

[0009] First, existing diffusion models primarily rely on data-driven learning during image restoration, lacking effective explicit structural prior constraints. When the input image is in an environment with extremely low signal-to-noise ratio, edge information and texture structure are easily submerged by noise, leading to problems such as structural artifacts, blurred contours, and loss of geometric details during diffusion sampling, making it difficult to achieve stable and reliable structural restoration.

[0010] Secondly, due to limitations in video memory and computing resources, high-resolution low-light images are typically trained and inferred using fixed-size image patches. Existing methods often focus only on local region features and lack the ability to model the global spatial relationships and scale information of the image. This leads to problems such as brightness discontinuities, color inconsistencies, and block artifacts in the enhancement results between different image patches, thus affecting the overall visual quality and spatial consistency.

[0011] Furthermore, while existing frequency domain enhancement methods incorporate frequency analysis tools such as Fourier transform or wavelet transform, they mostly focus on modeling amplitude spectrum information, neglecting the phase spectrum, which contains rich structural semantics and edge features. Since phase information plays a crucial role in image structure restoration, relying solely on amplitude spectrum modulation can easily lead to problems such as loss of fine-grained texture, blurring of structural details, and decreased reconstruction fidelity.

[0012] Therefore, how to coordinate and integrate prior structural information, global spatial positional relationships, and frequency domain phase features within the framework of a conditional diffusion model to achieve unified optimization of noise suppression, structure preservation, texture restoration, and spatial consistency is a core technical problem that urgently needs to be solved. Summary of the Invention

[0013] The purpose of this invention is to overcome the shortcomings of the prior art and provide a spatial-frequency-scale coordinated guided diffusion method for low-light image enhancement.

[0014] To achieve the objectives of this invention, the following technical solutions are adopted.

[0015] This invention provides a spatial-frequency-scale coordinated guided diffusion method for low-light image enhancement. The overall method flow is divided into two stages: a condition generation stage and a conditional diffusion recovery stage.

[0016] The condition generation stage is used to construct guiding information for the conditional diffusion network, including the generation of structural prior guiding conditions and the generation of global-scale location embedding features. Specifically, the generation of structural prior guiding conditions involves extracting a structural prior map from the input low-light image using the Sobel operator, extracting low-frequency components from the low-light image using wavelet decomposition, and inputting the structural prior map and low-frequency components into a two-stage wavelet structural prior guiding module (DWSPG). Structural prior guiding condition features are generated through cross-attention computation and multi-branch feature fusion. The generation of global-scale location embedding features utilizes a global-scale location embedding module (GSPE) to perform joint sinusoidal position encoding and nonlinear mapping on the current diffusion time step information, the spatial coordinates of the current image patch in the original image, and the original image cropping scale information to generate global-scale location embedding features. These features are then broadcast as global condition information for subsequent network computations.

[0017] The conditional diffusion recovery stage is used to progressively denoise and restore noisy images under the constraints of conditional guidance information. It includes an encoding stage, an intermediate feature refinement stage, and a decoding stage. In the encoding stage, the structural prior guidance features, global scale location embedding features, and the noisy image of the current diffusion step are jointly input into the conditional diffusion network. The spatial features are enhanced using the spatial attention module in the spatial feature extraction module, and a residual frequency-domain phase mixer (RFPM) is introduced to map the features to the Fourier frequency domain. The amplitude spectrum and phase spectrum are transformed and enhanced, and the phase residuals are corrected, respectively. After inverse transformation to restore to the spatial domain, spatial-frequency residual fusion is performed. In the intermediate feature refinement stage, the encoded deep features are subjected to bottleneck feature refinement processing. In the decoding stage, the bottleneck features are input into the multi-granularity guided reconstruction module. The decoded features are aggregated using multiple deep separable convolutional branches with different receptive fields and a gated feedforward network. The encoded features are then combined with skip connections to perform multi-granularity feature reconstruction. Finally, the predicted noise of the current diffusion time step is output, and the enhanced normal exposure image is output through back-diffusion denoising iteration.

[0018] A spatial-frequency-scale coordinated guided diffusion method for low-light image enhancement includes the following steps:

[0019] S1. Generate structural prior guidance conditions: Extract the structural prior map from the input low-light image using the Sobel operator, and simultaneously extract the low-frequency components by wavelet decomposition of the low-light image; input the structural prior map and the low-frequency components into the two-stage wavelet structural prior guidance module (DWSPG), and generate structural prior guidance condition features through cross-attention calculation and multi-branch feature fusion processing.

[0020] S2. Generate global scale location embedding features: Use the global scale location embedding module (GSPE) to perform joint sinusoidal location encoding and nonlinear mapping on the current diffusion time step information, the spatial coordinate information of the current image patch in the original image, and the cropping scale information of the original image to generate global scale location embedding features.

[0021] S3. Performing Conditional Diffusion Recovery and Feature Enhancement: The structural prior guiding conditional features, global scale location embedding features, and the noisy image of the current diffusion step are jointly input into the conditional diffusion network. In the encoding stage, the spatial attention module in the spatial feature extraction module is used to enhance the spatial features, and a residual frequency domain phase mixer (RFPM) is introduced to map the features to the Fourier frequency domain. The amplitude spectrum and phase spectrum are transformed and enhanced, and the phase residuals are corrected, respectively. After inverse transformation to restore to the spatial domain, spatial-frequency residual fusion is performed, and the encoded deep features are input into the intermediate feature refinement module for bottleneck feature refinement. The global scale location embedding features are injected into the encoding stage, intermediate feature refinement stage, and decoding stage through broadcasting to maintain global spatial consistency and scale consistency.

[0022] S4. Multi-granularity decoding and image reconstruction: The bottleneck features processed by the intermediate feature refinement module are passed to the multi-granularity guided reconstruction module in the decoding stage. The decoded features are aggregated using multiple depthwise separable convolutional branches with different receptive fields and gated feedforward networks, and multi-granularity feature reconstruction is performed by combining the encoded features passed by skip connections. The features output by the decoding stage are used as the prediction noise of the current diffusion time step, and the enhanced normal exposure image is output through back-diffusion denoising iteration.

[0023] Compared with the prior art, the present invention has the following advantages:

[0024] (1) Precise structural preservation

[0025] The two-stage wavelet structure prior guidance module (DWSPG) effectively suppresses noise interference and accurately preserves edge contours through structure guidance using low-frequency wavelet components. Experiments demonstrate that it achieves an SSIM of 0.930 on the LOLv2-Real dataset, exhibiting superior structure recovery performance compared to existing methods.

[0026] (2) Global consistency guarantee

[0027] The Global Scale Location Embedding (GSPE) module uses four-dimensional joint embedding to enable the network to perceive the global location and scale information of patches, effectively eliminating stitching artifacts and color inconsistencies during high-resolution image patch processing.

[0028] (3) Refined texture restoration

[0029] The Residual Frequency Domain Phase Mixer (RFPM) learns from phase residuals to finely recover high-frequency texture details while maintaining structural integrity. It achieves an LPIPS of 0.130 on the LOLv2-SYN dataset, demonstrating excellent visual perception quality.

[0030] (4) Collaborative optimization balance

[0031] The Space-Frequency Cooperative Feature Extraction (SFE Block) and the Multi-Granularity Guided Reconstruction (MGD Block) work together to achieve the optimal balance between noise reduction, detail preservation, and color fidelity, generating results with a visual naturalness close to that of real images. Attached Figure Description

[0032] Figure 1 This is a diagram of the overall architecture of the CS³-Diff framework. The model uses U-Net as the backbone network, comprising an encoder, intermediate layers, and a decoder. Red arrows indicate the main information flow. Given a low-light input image, a two-stage wavelet prior guidance module (DWSPG) extracts structure-aware conditions and injects them into the network via channel concatenation. The encoder employs stacked space-frequency collaborative feature extraction (SFE) blocks, integrating spatial attention (SpAM) and residual frequency-domain phase mixer (RFPM). The decoder utilizes a multi-granularity guided reconstruction (MGD) block for progressive reconstruction. An intermediate layer at the center of the network introduces a residual attention (REA) module for refining compact feature representations. Furthermore, a global scale location embedding (GSPE) module is broadcast to all stages to provide global context awareness; skip connections are used to preserve spatial details across resolutions.

[0033] Figure 2 This is an architecture diagram of the Two-Stage Wavelet Structure Prior Guidance (DWSPG) module; where:

[0034] (a) In the enhancement stage of DWSPG, the low-frequency component I_{LL} serves as keys and values, and the extracted structural prior serves as queries. This guides a cross-attention mechanism to achieve structure-aware feature enhancement. The enhanced features are further refined through residual blocks and fused with the low-frequency representation reconstructed by inverse discrete wavelet transform (IDWT) to generate the final enhanced features. (b) In the fusion stage of DWSPG, multiple parallel fusion branches aggregate the original low-light image I and the enhanced features F_{enhanced}. The outputs of each branch are concatenated and fed into a convolutional fusion head to generate the final conditional input S.

[0035] (c) Architecture of a single feature fusion branch. This module concatenates the original full-resolution image I with the enhanced features F_{enhanced}, and then uses a channel attention mechanism to adaptively weight the input features, thereby effectively fusing structural priors and rich texture details in the original image while suppressing noise;

[0036] Figure 3 This is an architecture diagram of the Global Scale Position Embedding Module (GSPE).

[0037] Figure 4 This is a structural diagram of the components inside the SFE module;

[0038] Figure 5 This is a structural diagram of the internal components of the MGD module;

[0039] Figure 6 The image shows the visual contrast results on the LOLv2-real low-light dataset.

[0040] Figure 7 A visualization comparison on the LOLv2-Synthetic dataset;

[0041] Figure 8 A visualization comparison chart on the LOLv1 dataset;

[0042] Figure 9 This is a visual comparison of object detection results in low-light (left) and enhanced-light (right) scenes on the ExDark dataset. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings and embodiments.

[0044] As an embodiment of the present invention, an image enhancement network CS³-Diff based on a structure-guided conditional diffusion model is provided. This network is based on conditional diffusion probabilistic models and achieves robust restoration from low-light input to high-quality normal-light images by introducing structure enhancement, spatiotemporal scale conditional embedding, and frequency domain feature modulation mechanisms.

[0045] I. CS³-Diff

[0046] The CS³-Diff framework proposed in this invention. For example... Figure 1 As shown, CS³-Diff is a generative framework based on a conditional diffusion model. Its core idea is to use multi-scale structural information extracted from low-light images as a condition. The framework guides the diffusion model to gradually recover a high-quality image with normal exposure during the reverse denoising process. It not only includes the forward denoising and reverse denoising processes of the standard diffusion model, but also innovatively introduces three key modules: DWSPG, GSPE, and RFPM. These modules are used for structural information extraction, spatiotemporal scale conditional embedding, and frequency domain feature enhancement, respectively, forming a clearly structured and perceptually robust enhancement system.

[0047] 1. Forward diffusion process

[0048] The forward diffusion process is defined as a fixed Markov chain that will produce a clear, properly exposed image. The gradual perturbation is Gaussian noise. Given... ,exist Gaussian noise is gradually injected in each diffusion step. At each time step... Above, the transition distribution is defined as:

[0049]

[0050] in This represents a predefined variance scheduling. Since the Gaussian distribution has a closed-form property, sampling at any time step t can be directly expressed as:

[0051]

[0052] in and As T increases, the final latent variable... Converges on an isotropic Gaussian distribution .

[0053] 2. Reverse diffusion process

[0054] The reverse diffusion process aims to extract from noisy samples Reconstruct a clean image Unlike the unconditional diffusion model, CS³-Diff formalizes the process as a conditional generation task, where the denoised trajectory is guided by auxiliary conditional information C.

[0055] Specifically, condition C integrates structural priors extracted by the DWSPG module, low-light image features, and global embeddings generated by GSPE. Therefore, the reverse process is parameterized as follows:

[0056]

[0057] in and These represent the predicted mean and variance, respectively.

[0058] Following common practice in diffusion models, the variance is fixed as:

[0059]

[0060] In this formalization, the network is trained to predict injected noise. Thus, the mean is obtained:

[0061]

[0062] The estimate of the clean image can be further obtained as follows:

[0063]

[0064] This method provides direct supervision during training.

[0065] 3 Reconstruction and Optimization Based on Backdiffusion

[0066] During the inference phase, sampling is performed from pure Gaussian noise. Begin by iteratively performing inverse denoising according to the following formula (7):

[0067]

[0068] in , Control the sampling variance.

[0069] To effectively optimize network parameters, we employ a hybrid loss function that jointly constrains the consistency of brightness and color, as well as the consistency of structure and perceptual fidelity.

[0070] To promote the reconstruction of loyalty, Loss of forced brightness and color consistency:

[0071]

[0072] In addition, a perceptual loss based on pre-trained VGG features is employed:

[0073]

[0074] in Indicates the activation of the l-th layer. The corresponding weights.

[0075] Furthermore, structural consistency is maintained through SSIM-based loss:

[0076]

[0077] The overall loss function is defined as:

[0078]

[0079] The diffusion loss is:

[0080]

[0081] Hyperparameters are empirically set to , and To balance reconstruction accuracy and perception fidelity.

[0082] II. Dual-stage Wavelet Structure Prior Guidance (DWSPG) module

[0083] Low-light images are often accompanied by severe noise interference, which weakens the reliability of pixel-domain structural information extraction. To address this issue, this paper proposes a two-stage wavelet structure prior guidance module (DWSPG). This module combines the decorrelation properties of wavelet transform in the frequency domain with the global modeling capability of attention mechanisms to effectively enhance and guide structural information.

[0084] Unlike traditional bi-branch wavelet structures that mainly rely on parallel multi-frequency feature extraction, the proposed DWSPG constructs a two-stage structure-guided framework for low-light enhancement tasks. Existing wavelet domain methods typically have two main problems: (1) high-frequency components are easily dominated by noise, resulting in blurred structural information; (2) there is a lack of effective mechanisms to re-inject reliable structural priors into the full-resolution image reconstruction process.

[0085] To address the aforementioned issues, DWSPG constructs a progressive structure-guided process. First, a structure-guided attention mechanism is used to extract noise-robust structural priors from low-frequency components. Then, the enhanced structural information is adaptively injected into the spatial domain reconstruction process, achieving detail restoration and artifact suppression. Compared to traditional wavelet eigenvalue decomposition methods, this design can more effectively maintain image structural consistency under extremely low signal-to-noise ratio conditions.

[0086] 1. Enhancement Phase:

[0087] Since high-frequency components are often overwhelmed by noise under low-light conditions, directly enhancing them can easily lead to noise amplification. In contrast, low-frequency components carry the main energy and structural information of the image and have stronger robustness. Based on this, the core strategy in this stage is to use two-dimensional discrete wavelet transform (DWT) to separate the signal and use only stable low-frequency components as a basis to introduce explicit structural priors for guided enhancement.

[0088] like Figure 2 As shown in Figure (a), given an input low-light image First, a one-layer two-dimensional discrete wavelet transform (DWT) based on the Haar wavelet basis is used to decompose the input image, obtaining low-frequency components and three high-frequency components:

[0089]

[0090] in This represents the discrete wavelet transform operator. Low-frequency components. It contains global illumination information and subject structure information of the image, while These correspond to high-frequency detail information in the horizontal, vertical, and diagonal directions, respectively.

[0091] To explicitly extract image structure features, the Sobel operator is first used to extract a structure prior map from the input image. Subsequently, a structure-guided cross-attention mechanism is constructed to generate keys and values ​​using low-frequency components, and to generate queries using the structure prior graph:

[0092]

[0093] in, All are implemented using 1×1 convolutions.

[0094] Then, structural enhancement features are obtained through cross-attention calculation:

[0095]

[0096] in, The channel dimension represents the key vector. This process utilizes explicit edge information to guide the low-frequency feature response, enabling the network to focus on structurally consistent regions while effectively suppressing background noise interference. The obtained attention-enhanced features are further processed by a nonlinear projection module and fused with the spatial domain low-frequency representation reconstructed through inverse wavelet transform to generate enhanced structural features:

[0097]

[0098] in, Represents the feature projection unit. This indicates the output convolution module.

[0099] After the enhancement phase, the network obtains a relatively clean global structural skeleton. However, since this phase mainly models low-frequency information, fine-grained texture details may be lost. Therefore, a further fusion phase needs to be designed to reinject the enhanced structural prior into the full-resolution spatial domain features.

[0100] 2. Fusion Stage

[0101] The fusion phase aims to integrate the enhanced structural prior features Adaptive fusion with the original input image achieves collaborative recovery of structural and texture information, such as... Figure 2 As shown in Figure (b).

[0102] First, the original input image is concatenated with the enhanced structural features:

[0103]

[0104] When the spatial resolutions of the two components are inconsistent, bilinear interpolation is used to achieve size alignment. Subsequently, multiple parallel fusion branches with different receptive fields are constructed, such as... Figure 2 As shown in Figure (c), each branch uses two convolutional layers for feature extraction:

[0105]

[0106] in, , Let represent the convolution kernel parameters in the (i)th fusion branch, with each branch having the same structure but independent parameters.

[0107] To achieve an adaptive balance between structural priors and original texture information, a channel attention mechanism is introduced to calculate the residual fusion weights. The output of each fusion branch can be represented as:

[0108]

[0109] in, Residual connections are generated using global average pooling and a multilayer perceptron (MLP). These connections not only preserve identity information from the input image but also improve the stability of the model training process.

[0110] Finally, the outputs of all fusion branches are aggregated to generate the final structural condition guidance information, which is then used as the condition input for the subsequent denoising process of the diffusion model, thereby providing reliable structural constraints for image restoration.

[0111]

[0112] Through this two-stage design, DWSPG not only achieves a balance between noise reduction and detail preservation, but more importantly, it becomes the main source of structural information throughout the diffusion process. The extracted structural priors are then propagated as conditional guides: these guides are used by GSPE for spatial alignment and by RFPM for frequency consistency reconstruction optimization, thus forming the basic framework of the proposed collaborative enhancement process.

[0113] III. Global-Scale-Position Embedding (GSPE)

[0114] In traditional diffusion models, the diffusion time step (t) is typically represented using sinusoidal positional embedding and injected into the network as conditional information to characterize the noise level in the current diffusion process. However, this type of conditional encoding, which relies solely on the time dimension, is essentially a one-dimensional conditional modeling approach. It implicitly assumes that all spatial locations in the image have a consistent feature processing method, thus ignoring the influence of differences in the spatial distribution of the image.

[0115] Unlike typical image generation tasks, low-light image enhancement exhibits significant spatial heterogeneity and scale dependence. Different spatial regions often possess different illumination distributions and texture structures, and the same texture may also display different frequency characteristics at different scales. Therefore, relying solely on diffusion time steps as conditional information is insufficient to fully characterize the spatial context relationships in complex low-light scenes.

[0116] To address the aforementioned issues, this paper proposes a global-scale-position embedding (GSPE) mechanism. For example... Figure 3 As shown, this method, based on the traditional diffusion time condition, further introduces global spatial location information and image scale information as auxiliary conditions, thereby enhancing the model's ability to perceive the global context.

[0117]

[0118]

[0119] Specifically, for each input image patch, GSPE extracts four conditional variables: the diffusion time step (t), the image patch row coordinates (i), the image patch column coordinates (j), and the image cropping scale (osize). These variables are then mapped to a high-dimensional embedding space using a sinusoidal positional coding function to obtain the temporal embedding. Row coordinate embedding Column coordinate embedding and scale embedding .

[0120]

[0121] Subsequently, the four embedding features are concatenated and input into a multilayer perceptron (MLP) with SiLU activation functions for feature fusion, thereby obtaining the final global scale location embedding representation. This embedding not only includes temporal information during the diffusion process but also integrates spatial location and scale information, providing the network with richer global contextual constraints.

[0122]

[0123] By introducing the GSPE mechanism, the model can perceive the real spatial location and scale relationship of the image patch in the original image during the image patch inference process, thereby effectively alleviating the problems of illumination discontinuity, block effect and insufficient global consistency caused by local inference in the high-resolution image enhancement process, and improving the spatial continuity and visual naturalness of the enhancement results.

[0124] IV. Space-Frequency Cooperative Feature Extraction Module (SFE Block)

[0125] While the U-Net architecture demonstrates powerful feature extraction capabilities, standard convolutional operations are inherently limited by local receptive fields, making it difficult to efficiently capture the global contextual information and frequency domain characteristics of images. To comprehensively enhance feature representation in low-light environments, we designed a hybrid processing paradigm—the SFE module—which operates simultaneously in the spatial and frequency domains within the encoder. The SFE module consists of a Spatial Attention Module (SpAM) and a proposed Residual Frequency-Phase Mixer (RFPM), which are cascaded together to work complementaryly and collaboratively, such as... Figure 4 As shown. This design aims to leverage the global perception characteristics of frequency domain representation to overcome the limitations of spatial domain feature extraction.

[0126] 1. Spatial Attention Module

[0127] To effectively capture local dynamic features and filter key structural information, this paper introduces the Spatial Attention Module (SpAM) from DarkIR. For example... Figure 4 As shown in Figure (a), this module aims to extract more effective feature representations through a feature selection mechanism in the spatial dimension.

[0128] For input features First, layer normalization (LN) is used to standardize the feature distribution to improve the stability of the training process. Then, 1×1 convolution and 3×3 convolution are used sequentially to extract multi-scale spatial feature information.

[0129] To achieve dynamic feature selection, a simple gating mechanism is introduced. Specifically, the convolutional output features are divided into two parts along the channel dimension. and Adaptive feature selection is achieved through element-wise multiplication.

[0130]

[0131] This gating mechanism can enhance the response in critical areas while suppressing redundant information.

[0132] Furthermore, to enhance the dependencies between spatial dimensions, a simplified channel attention (SCA) mechanism is introduced. This mechanism first extracts global statistics using global average pooling, then generates channel weights through linear transformation, and recalibrates the features. The final output features are represented as follows:

[0133]

[0134] in, This represents the learnable weight parameters. After SpAM processing, the network can obtain spatial domain feature representations with stronger local structure awareness.

[0135] 2. Residual Frequency-Phase Mixer (RFPM)

[0136] While SpAM can effectively model local context information, it is essentially a spatial domain operation and has limited ability to model global long-range dependencies. Therefore, this paper further designs a frequency domain enhancement module, RFPM, to achieve global frequency information interaction and structural detail recovery.

[0137] Existing frequency domain enhancement methods typically perform joint transformations on the Fourier amplitude and phase spectra. However, the phase spectrum contains rich image structural semantics and edge information. Directly applying complex nonlinear transformations to the phase spectrum can easily destroy the original spatial structure and generate artifacts during reconstruction. Furthermore, when the image itself does not require phase adjustment, forcing the learning of a complete phase map increases the difficulty of network optimization.

[0138] To address the aforementioned issues, this paper proposes a residual frequency domain phase mixer (RFPM). For example... Figure 4 As shown in Figure (b), this module introduces a residual guidance mechanism on the basis of full-spectrum interaction, enabling the network to learn phase residual corrections instead of regenerating the complete phase spectrum.

[0139] Specifically, for features after spatial enhancement First, the Fast Fourier Transform (FFT) is used to map it to the frequency domain and decompose it into an amplitude spectrum. and phase spectrum .

[0140] Subsequently, the amplitude spectrum and phase spectrum are input into two parallel multilayer perceptron branches for feature modeling:

[0141]

[0142] Each MLP consists of a 1×1 convolution and a LeakyReLU activation function, used to achieve frequency domain feature interaction.

[0143] Unlike traditional methods, RFPM does not directly predict the new phase spectrum. Instead, it models the phase branch output as a residual correction to the original phase. A scaling factor is introduced to control the magnitude of structural changes and improve training stability. :

[0144]

[0145] in, Used to limit the phase correction amplitude, thereby preventing excessive disturbance to structural information.

[0146] At the same time, the amplitude branch output is directly used as the new amplitude spectrum:

[0147]

[0148] This residual phase design is the core optimization strategy of RFPM. When the image structure does not need to be adjusted, the network only needs to learn to make the residual approach zero to retain the original phase information. This is easier to optimize than directly learning the complete phase mapping, and at the same time, it can effectively protect the fragile structural information in low-light images.

[0149] Finally, the corrected amplitude and phase spectra are recovered to the spatial domain using inverse fast Fourier transform (IFFT) to obtain the frequency domain enhancement features:

[0150]

[0151] Through this process, RFPM can fully extract global structural information in the frequency domain, improving texture detail recovery and reconstruction fidelity.

[0152] 3. Module integration and adaptive balancing mechanism

[0153] The frequency domain features extracted by RFPM are not used to replace spatial domain features, but rather as supplementary information re-injected into the main feature stream. Therefore, this paper designs learnable weight coefficients. This enables adaptive fusion of spatial and frequency domain information.

[0154] The final output can be represented as:

[0155]

[0156] in, These are learnable parameters, initialized to zero.

[0157] The zero-initialization strategy is one of the key design features of this module. In the early stages of training, due to... The network degenerates into a pure spatial attention-driven mode, thereby avoiding the unstable impact of the frequency domain branch on the training process. As training progresses, the network can adaptively learn appropriate weight values ​​based on the loss function, achieving a smooth transition from spatial domain optimization to joint spatial-frequency domain optimization.

[0158] This fusion mechanism not only overcomes the limitations of the local receptive field in the spatial domain, but also fully utilizes the structural prior information provided by DWSPG and the global spatial consistency constraints provided by GSPE. Since the input features processed by RFPM already possess structural awareness and spatial consistency, they can focus more on fine-grained frequency domain texture optimization without having to undertake the task of global structural correction.

[0159] Therefore, in the entire CS³-Diff framework, RFPM, as the final frequency domain enhancement stage, works in concert with phase-aware residual modulation mechanism, structure prior guidance, and spatial location constraints to achieve high-fidelity texture reconstruction and structure preservation, thereby significantly improving the enhancement quality of low-light images.

[0160] 5. Multi-Granularity Guided Reconstruction Module (MGD Block)

[0161] This module mainly consists of an inflated spatial attention module (Di-SpAM) and a gated feedforward network (Gated-FFN), such as Figure 5 As shown, Di-SpAM employs a parallel dilated convolution structure to acquire multi-scale receptive field information without significantly increasing the number of parameters, thereby simultaneously modeling local texture details and global illumination distribution, improving feature representation capabilities. Gated-FFN uses a gating mechanism to achieve channel-dimensional feature selection and nonlinear reparameterization, enhancing the response of important features and suppressing background noise interference. Simultaneously, a residual connection mechanism ensures stable gradient propagation.

[0162] By combining the spatial context modeling capabilities of Di-SpAM with the channel feature filtering capabilities of Gated-FFN, MGDBlock can effectively fuse features at different scales, improving image contrast restoration, texture reconstruction, and noise suppression capabilities, and providing richer feature representations for subsequent image reconstruction.

[0163] V. Experiment

[0164] 1. Dataset

[0165] To fully verify the effectiveness of the proposed CS³-Diff model in low-light image enhancement tasks, this paper selects four publicly available and widely used low-light image enhancement benchmark datasets for experimental evaluation, including LOLv1, LOLv2-Real, LOLv2-Synthetic, and LSRW datasets.

[0166] (1) LOLv1 dataset

[0167] LOL-v1 is one of the most representative public datasets in the field of low-light image enhancement. It consists of paired low-light images captured in real-world scenes and their corresponding normally exposed images. This dataset contains 485 pairs of training images and 15 pairs of test images, and is widely used for performance evaluation of low-light image enhancement methods.

[0168] (2) LOLv2-Real dataset

[0169] LOLv2-Real is a real-world low-light dataset that extends LOL-v1, containing a wider variety of scene types and lighting conditions. Following the official partitioning method, it includes 689 pairs of training images and 100 pairs of test images to evaluate the model's enhancement capabilities in real-world low-light environments.

[0170] (3) LOLv2-Synthetic dataset

[0171] LOLv2-Synthetic is a synthetic low-light dataset that simulates low-light degradation by adjusting the luminance channel (Y channel) of RAW images. This dataset contains 900 pairs of training images and 100 pairs of test images, which can be used to validate the model's generalization performance in synthetic degradation scenarios.

[0172] (4) LSRW dataset

[0173] LSRW is a large-scale real-world low-light image augmentation dataset, with images acquired using Nikon SLR cameras and Huawei smartphones. The Nikon subset contains 3150 training image pairs and 20 test image pairs, while the Huawei subset contains 2450 training image pairs and 30 test image pairs. Because this dataset covers a wide range of real-world scenes and shooting devices, it can effectively evaluate the augmentation performance and generalization ability of models in complex real-world environments.

[0174] 4.2 Experimental Setup

[0175] This paper implements the model based on the PyTorch framework and trains it using the Adam optimizer. To enhance the model's adaptability to features at different spatial scales and to verify the effectiveness of the GSPE module, a multi-scale training strategy is adopted during the training phase. Specifically, the input image is randomly cropped into image patches of different scales, such as 64×64, 128×128, and 256×256, and then uniformly scaled to 64×64 as the network input. The original cropping scale is recorded as scale condition information in the GSPE module, thereby enhancing the model's ability to perceive scenes at different spatial scales. A linear noise scheduling strategy is used during the diffusion process, with a total of 1000 diffusion steps.

[0176] V. Evaluation Indicators

[0177] To objectively evaluate the enhancement effect of the proposed method, peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and learned perceptual image patch similarity (LPIPS) are used as evaluation metrics.

[0178] Among them, PSNR is used to measure the pixel-level reconstruction error between the enhanced image and the reference image; SSIM is used to evaluate the image's ability to preserve brightness, contrast and structural information; LPIPS uses deep features to calculate perceptual similarity, which reflects how close the enhancement result is to the real image at the human eye level.

[0179] Generally speaking, the higher the PSNR and SSIM values ​​and the lower the LPIPS value, the closer the enhancement result is to the reference image and the better the enhancement effect.

[0180] Table 0. Summary of Quantitative Results:

[0181] Dataset PSNR↑ SSIM↑ LPIPS↓ LOLv2-Real 25.12 0.93 0.14 LOLv2-Synthetic 23.5 0.95 0.13 LOLv1 24.26 0.9 0.22 LSRW 23.25 0.715 0.32

[0182] Table 1. Quantitative comparison results on the LOLv2-Real dataset.

[0183] Methods PSNR↑ SSIM↑ LPIPS↓ RetinexNet 15.68 0.782 0.317 DRBN 20.13 0.822 0.135 RUAS 15.35 0.670 0.375 Zero-DCE 18.05 0.602 0.362 SNR-SKF 20.66 0.824 0.191 LLDiffusion 23.16 0.842 / Bread 20.83 0.841 0.234 DiffLL 19.61 0.803 0.210 LLFormer 21.50 0.540 0.308 FourLLIE 22.34 0.833 0.106 RetinexFormer 22.79 0.761 0.273 ZMAR-SNFlow 24.15 0.862 0.138 RetinexMamba 27.31 0.832 0.172 SG-LLIE 22.84 0.859 0.126 CIDNet 24.11 0.871 0.108 JTE-CFlow 24.06 0.878 / MobileIE 25.08 0.845 0.180 UPT-Flow 25.06 0.890 0.115 GAMP 24.00 0.845 / Ours 25.12 0.930 0.140

[0184] Table 2. Quantitative comparison results on the LOLv2-Synthetic dataset.

[0185] Methods PSNR↑ SSIM↑ LPIPS↓ RetinexNet 17.14 0.756 0.255 Zero-DCE 17.76 0.814 0.168 MIRNet 21.94 0.876 0.317 RUAS 13.4 0.640 0.364 KinD++ 19.26 0.803 0.226 Restormer 21.41 0.843 0.102 URetinex-Net 18.75 0.826 0.192 CDAN 22.35 0.850 0.150 SCI 15.43 0.744 0.233 SNR-SKF 17.21 0.774 0.073 Pydiff 17.33 0.797 0.255 LLFormer 17.16 0.784 0.244 Bread 17.63 0.919 0.091 DPFNet 23.9 0.840 0.130 LLDiffusion 25.99 0.948 / ZMAR-SNFlow 25.62 0.928 0.037 JTE-CFlow 25.21 0.942 / UPT-Flow 26.48 0.946 0.037 GAMP 26.11 0.940 / Ours 23.5 0.950 0.130

[0186] Table 3 shows the quantitative comparison results on the LOLv1 dataset.

[0187] Methods PSNR↑ SSIM↑ LPIPS↓ RetinexNet 16.77 0.463 0.475 KinD 20.87 0.800 0.207 ZeroDCE 14.86 0.559 0.335 AnlightenDiff 21.72 0.814 0.141 LLFormer 23.64 0.819 0.169 RUAS 16.4 0.503 0.364 IAT 21.25 0.844 0.255 FourLLIE 20.99 0.807 0.095 SNR-SKF 20.51 0.711 0.181 Retinexformer 22.71 0.818 0.092 LightenDiffusion 20.19 0.815 0.316 HWMNet 19.62 0.862 0.271 RetinexMamba 23.15 0.821 0.088 DMFourLLIE 22.98 0.827 0.079 DiffLL 26.33 0.845 0.217 SG-LLIE 24.63 0.873 0.092 MobileIE 23.62 0.812 0.198 RIRO 23.01 0.897 0.072 GAMP 24.96 0.846 / Ours 24.26 0.900 0.220

[0188] Table 4 shows the quantitative comparison results on the LSRW dataset.

[0189] Methods PSNR↑ SSIM↑ LPIPS↓ RetinexNet 15.61 0.414 0.454 DRBN 16.73 0.507 0.457 RUAS 13.02 0.359 0.379 Restormer 16.30 0.453 0.427 Uformer 16.59 0.494 0.435 MIRNet 16.47 0.477 0.430 SNR-Aware 16.49 0.505 0.419 LLFormer 20.69 0.560 0.518 FourLLIE 18.61 0.505 0.316 SDTL 21.23 0.576 0.392 Retinexformer 20.15 0.534 0.336 DiffLL 17.27 0.397 0.439 GSAD 17.29 0.510 0.290 DMFourLLIE 21.47 0.623 0.189 PyDiff 17.01 0.517 0.343 WeatherDiff 15.58 0.486 LightenDiffusion 18.48 0.540 0.310 RIRO 19.96 0.744 0.228 SPNet 16.63 0.520 / Ours 23.25 0.715 0.320

[0190] To verify the enhancement performance of the proposed CS³-Diff model, experiments were conducted on the LOLv1, LOLv2-Real, LOLv2-Synthetic, and LSRW datasets, and PSNR, SSIM, and LPIPS were used as evaluation metrics. The experimental results are shown in Tables 0-4.

[0191] Experimental results show that CS³-Diff achieves good performance on multiple datasets. On the LOLv2-Real dataset, the model achieves a PSNR of 25.12 dB and a SSIM of 0.93, indicating that the proposed method can effectively recover image brightness information and maintain high structural consistency. On the LOLv2-Synthetic dataset, the model achieves an SSIM of 0.95 and LPIPS of 0.13, demonstrating that the proposed structure prior guidance and frequency domain enhancement mechanism can effectively recover image structure and visual details. On the LOLv1 dataset, CS³-Diff achieves a PSNR of 24.26 dB and an SSIM of 0.90, further validating the model's stable enhancement capability under different low-light scenes. For the LSRW dataset, which has greater scene complexity and stronger realistic degradation characteristics, the model still achieves a PSNR of 23.25 dB and an SSIM of 0.715, indicating that the proposed method has good generalization ability and robustness.

[0192] In summary, CS³-Diff achieves a good balance between brightness restoration, structure preservation, and perceptual quality. This is because the DWSPG module provides reliable structural prior constraints, GSPE enhances the model's perception of global spatial context, and RFPM fully utilizes frequency domain phase information for fine-grained texture reconstruction, thereby effectively improving the enhancement quality of low-light images.

[0193] To further verify the visual enhancement effect of the proposed method, qualitative comparative experiments were conducted on the LOLv2-Real, LOLv2-Synthetic, and LOLv1 datasets. The experimental results are as follows: Figure 6-8As shown in the figure, on the LOLv2-Real dataset, some existing methods suffer from color distortion, noise residue, and structural blurring, while CS³-Diff can maintain a more natural color representation while improving brightness and effectively recovering details in slender structural areas such as railings. On the LOLv2-Synthetic dataset, although existing methods can improve overall exposure, they tend to over-enhance or lose details in complex architectural texture areas. In contrast, CS³-Diff can better recover architectural edges and texture structures, achieving a better balance between brightness recovery and structure preservation. On the LOLv1 dataset, the bookshelf scene contains a large amount of fine-grained texture information and complex lighting areas. Experimental results show that CS³-Diff can effectively suppress noise while preserving high-frequency details such as spine edges and wood textures, achieving a more natural and clear enhancement effect.

[0194] In summary, the proposed method demonstrates good performance in brightness recovery, color preservation, and structural detail reconstruction, validating the effectiveness of the synergistic effect of DWSPG, GSPE, and RFPM.

[0195] VI. Ablation Experiment

[0196] 1. Ablation of core modules

[0197] To verify the effectiveness of each key module in CS³-Diff, quantitative ablation experiments were conducted on the LOLv2-Real dataset. Specifically, ablation analysis was performed on the Dual-Stage Wavelet Structural Prior Guidance (DWSPG), Global-Scale Positional Embedding (GSPE), and Residual Frequency-domain Phase Mixer (RFPM). The experimental results are shown in Table 5.

[0198] Table 5 Ablation experiments conducted on the LOLv2-Real dataset to assess the effectiveness of the CS³-Diff key module.

[0199]

[0200] First, we analyze the contribution of the DWSPG module. As shown in the table above, the model performance deteriorates most significantly after removing DWSPG. Specifically, PSNR decreases from 25.12 dB to 23.87 dB, while LPIPS increases from 0.140 to 0.169. This result indicates that introducing wavelet domain structural priors is crucial for low-light image restoration. By decomposing the image into high-frequency and low-frequency components and utilizing edge-aware structural information to guide the diffusion denoising process, DWSPG can more effectively suppress noise and reconstruct details under extremely low signal-to-noise ratio conditions, thereby improving the overall image restoration quality.

[0201] The role of the GSPE module was then analyzed. Experimental results show that after removing GSPE, the model's SSIM decreased to 0.894, the lowest among all ablation settings, while PSNR only decreased slightly. This indicates that relying solely on local image patches for enhancement is insufficient to guarantee the global consistency of high-resolution images, easily leading to spatial reconstruction discontinuities. By introducing global scale location encoding, GSPE can provide the model with richer global contextual information, enabling it to simultaneously perceive the positional relationships and scale information of image patches during the reconstruction process, thereby effectively maintaining the continuity of the overall illumination distribution and structural consistency of the image.

[0202] Finally, the importance of the RFPM module is analyzed. Compared with the complete model, although the model without RFPM achieved a similar PSNR (25.07 dB), its perceptual quality significantly decreased, with LPIPS increasing to 0.169 and SSIM decreasing to 0.897. This phenomenon indicates that relying solely on spatial domain feature modeling is insufficient to fully recover high-frequency texture details in images. RFPM, by explicitly modeling and modulating phase information in the frequency domain, effectively enhances the sharpness and visual realism of texture edges, providing a beneficial supplement to amplitude spectrum-based information reconstruction, thereby further improving the structural fidelity and perceptual quality of the image.

[0203] In summary, the ablation experiments fully validated the effectiveness of the three core modules: DWSPG, GSPE, and RFPM. These three modules improve model performance from different angles, including structural prior guidance, global spatial awareness, and frequency domain texture enhancement. Their synergistic effect enables CS³-Diff to achieve a good balance between pixel-level reconstruction accuracy, structural consistency, and visual perception quality.

[0204] 2. DWSPG module ablation

[0205] To verify the necessity and complementary effect of the two-stage structure design in the DWSPG module, this paper further removed the fusion stage and retained only the enhancement stage for experiments. The results are shown in Table 6. Table 6: Ablation Experiment of Internal Mechanism of DWSPG Module:

[0206] DWSPG w / o Fusion 25.18 0.83 0.32 Ours 25.12 0.93 0.14

[0207] The experimental results show that removing the fusion stage slightly improved the model's PSNR from 25.12 dB to 25.18 dB, but significantly reduced the SSIM and LPIPS metrics. Specifically, SSIM decreased from 0.93 to 0.83, while LPIPS increased from 0.14 to 0.32. This phenomenon indicates that while relying solely on the frequency domain enhancement stage can reduce pixel-level errors and achieve a higher PSNR, it leads to excessive smoothing of image textures, thereby compromising structural consistency and visual perception quality.

[0208] The reason for this is that the enhancement stage mainly utilizes the low-frequency information after wavelet decomposition for structural enhancement and noise suppression, which can build a stable structural skeleton and improve the overall reconstruction accuracy. However, since this process focuses on low-frequency structural modeling, high-frequency texture details are easily weakened, resulting in an over-smoothing phenomenon in the generated result. Although this smoothing effect can reduce pixel-level errors, it also loses a lot of detail information that is more important to human visual perception.

[0209] In contrast, the fusion stage achieves the coordinated recovery of structural and texture information by adaptively fusing the enhanced structural features with the original input image in the spatial domain. On the one hand, this stage can preserve high-frequency texture details in the original image; on the other hand, it can make full use of the prior structural information extracted in the enhancement stage, thus achieving a better balance between noise suppression and detail preservation.

[0210] The table above further illustrates that after introducing the fusion stage, the model's SSIM improved from 0.83 to 0.93, while LPIPS decreased from 0.32 to 0.14, indicating that the fusion stage plays a crucial role in improving structure preservation and perceptual quality. Although the PSNR of the complete model is slightly lower than that of the model that only retains the enhancement stage, it shows significant advantages in both structural consistency and visual realism.

[0211] In summary, this ablation experiment validates the rationality of the two-stage design of DWSPG. The enhancement stage is responsible for building stable structural priors and suppressing noise, while the fusion stage further restores high-frequency texture details and enhances perceptual quality. Together, they achieve an effective balance between structure preservation, noise suppression, and detail reconstruction during low-light image enhancement.

[0212] 3. GSPE module ablation

[0213] To verify the effectiveness of the proposed Global-Scale Position Embedding (GSPE) in mitigating the spatial inconsistency problem in diffusion models, this paper further conducts progressive ablation experiments. By gradually increasing the conditional embedding dimension, the contribution of each component is analyzed. The experimental results are shown in Figure 7.

[0214] Table 7: Ablation experiment of multidimensional conditional embedding strategy for GSPE module.

[0215] t 25.07 0.85 0.24 t、(i,j) 25.10 0.89 0.17 Ours 25.12 0.93 0.14

[0216] First, embedding using only diffusion time steps As a baseline model, experimental results show that this model performs the worst, with PSNR, SSIM, and LPIPS of 25.07 dB, 0.85, and 0.24, respectively. This result demonstrates the significant limitations of relying solely on temporal conditional information. Traditional diffusion models assume that different spatial locations share the same feature processing methods, implicitly assuming spatial invariance in images, thus failing to adequately model the complex spatial distribution features in low-light images. During high-resolution image patch inference, this lack of spatial awareness easily leads to inconsistent enhancement in local regions, resulting in problems such as illumination discontinuities or block artifacts.

[0217] Subsequently, spatial coordinate information was further introduced based on the time step embedding. Experimental results show that the model performance is significantly improved, with SSIM increasing from 0.85 to 0.89 and LPIPS decreasing from 0.24 to 0.17. These results demonstrate that explicitly introducing spatial location information can effectively break the spatial invariance assumption of traditional diffusion models, enabling the model to perceive the specific location of image patches within the original image. This allows for better adaptation to local illumination variations and structural differences in different regions, improving spatial consistency and structural recovery capabilities.

[0218] Finally, scale information is further introduced based on time step and spatial coordinate encoding. To form a complete GSPE embedding strategy At this point, the model achieved the best results across all evaluation metrics, with PSNR reaching 25.12 dB, SSIM improving to 0.93, and LPIPS decreasing to 0.14. These results demonstrate that scale information can provide the model with additional global contextual constraints, enabling it to distinguish the relationship between local textures and global structure, thereby better recovering fine-grained texture details while maintaining overall structural consistency.

[0219] In summary, the progressive ablation experiments fully validated the effectiveness of each component of GSPE. Spatial coordinate encoding enhanced the model's ability to perceive spatial relationships, while scale encoding further provided global contextual information. Together with diffusion time-step encoding, these two components enabled the model to possess stronger spatial consistency modeling capabilities and adaptive enhancement capabilities, thereby effectively mitigating the spatial discontinuity problem generated during high-resolution image patch inference and significantly improving the structural fidelity and visual quality of the enhancement results.

[0220] 4. RFPM module ablation

[0221] To verify the importance of the phase modeling mechanism in the Residual Frequency-domain Phase Mixer (RFPM), this paper further conducts phase branch ablation experiments. Specifically, while keeping the amplitude spectrum branch unchanged, the phase residual learning branch (denoted as "RFPM w / o Phase") is removed, that is, only the amplitude spectrum is modulated while the original phase spectrum remains unchanged. The experimental results are shown in Table 8.

[0222] Table 8: Experimental verification of the effectiveness of the phase branch of the residual frequency domain phase mixer:

[0223] RFPM w / o Phase 24.84 0.9 0.15 Ours 25.12 0.93 0.14

[0224] As shown in Table 8, the experimental results reveal a significant decrease in model performance after removing the phase residual learning branch. Specifically, PSNR decreased from 25.12 dB to 24.84 dB, SSIM decreased from 0.930 to 0.900, and LPIPS increased from 0.140 to 0.150. These results demonstrate that phase information plays a crucial role in low-light image enhancement tasks, and relying solely on amplitude spectrum modeling is insufficient to fully recover structural details and texture information in images.

[0225] The amplitude spectrum in the Fourier transform mainly reflects the overall energy distribution and illumination information of an image, while the phase spectrum contains a large amount of structural semantic information, edge information, and spatial layout information. Many existing frequency domain enhancement methods mainly focus on amplitude spectrum modulation, while neglecting the important role of the phase spectrum, resulting in insufficient high-frequency detail recovery, and even problems such as texture blurring and structural misalignment.

[0226] In contrast, the RFPM proposed in this paper does not directly predict the new complete phase spectrum, but instead makes minor corrections to the original phase through a residual learning mechanism. Specifically, the network only learns sparse phase residuals and superimposes them as compensation terms onto the original phase spectrum. This design not only reduces the optimization difficulty but also enables targeted enhancement of fine-grained textures while maintaining the stability of the original structural information.

[0227] Experimental results show that after introducing phase residual guidance, the model can more accurately recover the edge structure and high-frequency texture details in the image, while maintaining the consistency of the global spatial structure, thus achieving better performance in terms of structural fidelity and visual perception quality.

[0228] In summary, this ablation experiment fully validates the effectiveness of the phase residual learning mechanism in RFPM. By jointly utilizing amplitude and phase spectrum information, the model can more fully exploit frequency domain structural features, achieving more realistic and high-fidelity low-light image enhancement results.

[0229] 5. Target Detection

[0230] To further evaluate the effectiveness of our method in downstream tasks, we conducted object detection experiments on the ExDark dataset. We used the standard detector YOLOv3 and report the mean accuracy (AP) and mean mean accuracy (mAP) for each class. The baseline represents the detection performance on the original low-light images without any enhancement.

[0231] Table 9. Quantitative results of low-light image target detection on the ExDark dataset.

[0232]

[0233] As shown in Table 9, existing image augmentation methods offer only limited or inconsistent performance improvements compared to the baseline. For example, KinD and Zero-DCE perform close to the baseline overall, but even show performance degradation in some categories, indicating that simple augmentation strategies are not always beneficial for detection tasks. While MEBBLN has achieved some improvement, it lacks consistency across different target categories.

[0234] In comparison, our method achieves the best overall performance, improving the mean AP from 72.0 to 73.6. Notably, our method delivers significant gains on challenging categories such as Boat and Bus, while maintaining stable performance on other categories. These results demonstrate that our method can effectively enhance low-light images in a structure-preserving manner, thus benefiting downstream object detection tasks.

[0235] like Figure 9As shown, this image presents a visual comparison of target detection results before and after enhancement in low-light scenes. In the original low-light scene, due to insufficient illumination and poor visibility, the detector cannot identify multiple targets or only provides low-confidence predictions. After enhancement using our method, more targets are successfully detected, and the confidence level is significantly improved, especially in challenging areas such as distant pedestrians and partially occluded targets.

[0236] 5. Conclusion

[0237] This paper proposes CS³-Diff, a unified diffusion framework that addresses key challenges in low-light image enhancement, including severe noise, structural degradation, and spatial inconsistencies, by introducing a collaborative spatio-frequency-scale guidance mechanism. Unlike methods that rely on isolated module designs, our model explicitly integrates structural priors, global context modeling, and frequency domain refinement into a consistent conditional generation process.

[0238] Extensive experiments on multiple benchmark datasets demonstrate that CS³-Diff exhibits consistent improvements in both structure fidelity and perceptual quality, specifically superior SSIM and competitive LPIPS performance. These results highlight the importance of structure-oriented optimization, especially in scenarios with extremely low signal-to-noise ratios, where relying solely on pixel-level metrics is often insufficient.

[0239] Despite the encouraging results achieved, this study still has several potential limitations. First, the high computational complexity of the model due to the diffusion model framework and multi-component design may limit its real-time application capabilities. Second, although the model demonstrates good generalization ability on the LSRW dataset, its performance may degrade when faced with unseen sensor characteristics or extreme lighting conditions. Third, the validity of the structural prior depends on the assumption that meaningful structures can be reliably extracted, and this assumption does not always hold true in severely degraded scenarios.

[0240] Despite these limitations, these problems are somewhat prevalent in the challenging task of low-light image enhancement. The continuous improvement in structural consistency and visual realism in the experiments demonstrates that the proposed method achieves a relatively ideal balance between model complexity and restoration quality. Future work will focus on further improving efficiency and robustness through model acceleration and adaptive generalization strategies.

Claims

1. A spatial-frequency-scale coordinated guided diffusion method for low-light image enhancement, characterized in that, Includes the following steps: S1. Simultaneously perform structural edge extraction and wavelet decomposition on the input low-light image. Send the obtained structural prior map and low-frequency components of the image into the two-stage wavelet structural prior guidance module. After cross-attention calculation and multi-branch feature fusion processing, generate structural prior guidance condition features. S2. Through the global scale location embedding module, joint sinusoidal location coding and nonlinear mapping are performed on the diffusion time step, image patch spatial coordinates, and image cropping scale to generate global scale location embedding features. S3. The structural prior guidance features, global scale location embedding features, and the noisy image of the current diffusion step are input into the conditional diffusion network. In the encoding stage, a spatial attention module is used to enhance spatial features. The features are mapped to the Fourier frequency domain through a residual frequency domain phase mixer. Transformation enhancement and phase residual correction are performed on the amplitude spectrum and phase spectrum respectively. After inverse transformation and propagation back to the spatial domain, spatial-frequency residual fusion is completed. The deep features of the encoding are input into the intermediate feature refinement module to achieve bottleneck feature refinement. At the same time, the global scale location embedding features are injected into the entire stage of encoding, feature refinement, and decoding in a broadcast manner to ensure the global spatial and scale consistency of the image. S4. The refined bottleneck features are input into the multi-granularity guided reconstruction module in the decoding stage. The decoded features are aggregated through multi-receptive field depth separable convolutional branches and gated feedforward networks. The encoded features are combined with the skip connection transmission to complete the multi-granularity feature reconstruction. The decoded output features are used as the prediction noise for the current diffusion time step. After back-diffusion denoising iteration, the enhanced image with normal exposure is finally output.

2. The spatial-frequency-scale coordinated guided diffusion method for low-light image enhancement according to claim 1, characterized in that, The dual-stage wavelet structure prior guidance module includes an enhancement stage sub-module and a fusion stage sub-module arranged sequentially. The enhancement stage sub-module includes a discrete wavelet transform unit and a structure graph extraction unit. The discrete wavelet transform unit uses the Haar wavelet basis to perform a two-dimensional discrete wavelet transform on the low-light image to extract low-frequency components, and uses convolutional projection as key and value features. At the same time, the structure graph extraction unit extracts the structure prior graph through the Sobel operator and uses convolutional projection as query features. The structure enhancement features are obtained through the cross-attention mechanism. After performing nonlinear projection and inverse discrete wavelet transform, the spatial domain enhanced structure features are obtained. The fusion stage submodule includes multiple sets of parallel feature fusion branches and convolutional fusion heads. Each branch stitches the low-light image with the enhanced structural feature channel, and completes single-branch feature fusion through convolution and channel attention. The convolutional fusion head concatenates all branch fusion features with the low-light image channels again, and outputs the structural prior guiding condition features after convolutional fusion.

3. The spatial-frequency-scale coordinated guided diffusion method for low-light image enhancement according to claim 1, characterized in that, The processing flow within the global-scale location embedding module includes: Using four parallel sinusoidal position encoding units, the received current diffusion time step information, the row coordinate information, column coordinate information of the current image patch in the original image, and the original image cropping scale information are sinusoidally encoded and mapped to the embedding space of a preset dimension to obtain the time step embedding vector, row position embedding vector, column position embedding vector, and scale embedding vector. The time step embedding vector, row position embedding vector, column position embedding vector, and scale embedding vector are concatenated along the channel dimension and input into a fusion multilayer perceptron for nonlinear mapping and feature fusion, outputting the global scale position embedding feature.

4. The spatial-frequency-scale coordinated guided diffusion method for low-light image enhancement according to claim 1, characterized in that, The internal processing flow of the residual frequency domain phase mixer includes: The input spatial features are transformed into the Fourier frequency domain using the Fast Fourier Transform, and the amplitude spectrum and phase spectrum are decoupled and separated. The amplitude spectrum is input into the first transform branch and then transformed and enhanced by two 1×1 convolutions and the LeakyReLU activation function to obtain the enhanced amplitude spectrum. The phase spectrum is input into the second transform branch, and the phase residual is predicted by passing two 1×1 convolutions and the LeakyReLU activation function in sequence. The phase residual is then multiplied by a scaling factor and added to the original phase spectrum to obtain the corrected phase spectrum. The enhanced amplitude spectrum and the corrected phase spectrum are recombined to form a complex spectrum, and the inverse fast Fourier transform is used to recover it to the spatial domain to obtain the frequency domain enhancement features. The frequency domain enhancement feature is multiplied by a learnable weighting coefficient initialized to zero, and then added to the spatial feature output by the spatial attention module as a residual to achieve synergistic enhancement of spatial domain features and frequency domain features.

5. The spatial-frequency-scale coordinated guided diffusion method for low-light image enhancement according to claim 1, characterized in that, The conditional diffusion network is trained using a hybrid loss function, which includes diffusion noise prediction loss, reconstruction loss, perception loss, and structural similarity loss. Each loss term is weighted and summed using preset weight coefficients. The hybrid loss function is expressed as follows: ； in, To predict the loss for diffused noise, for Reconstruction losses, In order to perceive loss, For structural similarity loss, These are the weighting coefficients for the corresponding loss terms.

6. The spatial-frequency-scale coordinated guided diffusion method for low-light image enhancement according to claim 1, characterized in that, The method employs a multi-scale training strategy during the training phase, randomly cropping the input image into image patches of three sizes: 64×64, 128×128, and 256×256, and uniformly scaling them to 64×64 before inputting them into the conditional diffusion network for training. The side length of the original cropped image patch is defined as a scale-aware parameter, and this scale-aware parameter is used as one of the inputs to the global scale location embedding module in joint location encoding. During the full-image inference phase, the scale-aware parameter is dynamically determined based on the resolution of the pre-processed and scaled input image to achieve scale-consistent feature extraction and image enhancement.

Images