Dynamic residual error mixing expert low-illumination image enhancement method and device and storage medium

Abstract

The application discloses a dynamic residual error mixed expert low-illumination image enhancement method and device and a storage medium, and relates to the technical field of image processing. The method separates and models image structure information and texture information in a latent representation space. Secondly, a mixed expert network and a multi-scale hierarchical denoising network structure are used to realize fine denoising and reconstruction. Finally, a unified model framework is constructed to restore key structures and detail features in low-light images. The application can be widely applied to image enhancement processing in the fields of night monitoring, automatic driving, mobile photography, remote sensing detection and the like.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and specifically to a dynamic residual hybrid expert low-light image enhancement method, apparatus, device, and storage medium. Background Technology

[0002] Low-light images refer to images acquired in low-light environments such as at night, in dimly lit indoor environments, or in tunnels. Due to insufficient light, low-light images generally suffer from low brightness, poor contrast, loss of detail information, and significant noise, which seriously affects the performance of subsequent visual tasks such as target detection, recognition, and segmentation.

[0003] Traditional low-light image enhancement methods mainly rely on histogram equalization, Retinex theory, deep learning-based methods, or filtering-based image enhancement techniques to improve visual effects by making global or local adjustments to image brightness or contrast. For example:

[0004] (1) Histogram equalization: By redistributing the statistical distribution of image gray values, pixels that were originally concentrated in dark areas are stretched to a wider brightness range, thereby improving the overall contrast and making details in dark areas easier to observe. It is simple to implement and has low computational cost, so it is often used as a basic enhancement method. However, it also has obvious limitations: First, it is easy to enhance noise as well, resulting in increased graininess in the image; second, it only focuses on the global gray distribution, which may cause over-enhancement or brightness distortion in local areas; third, if not handled properly in color images, it may also cause color deviation.

[0005] (2) Retinex-based method: Retinex decomposes an image into a reflection component that reflects the inherent properties of the object and an illumination component that is affected by illumination. Under low light conditions, the image darkens mainly due to insufficient illumination component. Therefore, brightness and detail can be improved by estimating and correcting the illumination component and enhancing the reflection component, while maintaining the naturalness and color stability of the image as much as possible. Compared with the traditional histogram method, Retinex is more in line with the human visual mechanism and performs better in non-uniform lighting scenes. However, its limitations are also obvious: first, the model solution is complex and the computation is large; second, it is sensitive to parameters and has limited generalization ability in different scenes; and third, it is prone to problems such as halo, color distortion or noise amplification.

[0006] (3) Filter-based methods: These methods improve visual effects by separating and adjusting different frequency components of the image. Generally, smoothing filters are used to estimate low-frequency illumination information in the image, and then high-frequency enhancement is used to highlight details such as edges and textures, thereby improving the visibility of dark areas and overall contrast. Some methods also combine homomorphic filtering to enhance reflective components while suppressing uneven illumination. These methods are relatively simple to implement and computationally efficient, making them suitable for basic image preprocessing. However, their limitations are also significant: the selection of filter parameters is sensitive, and the effects vary greatly in different scenes; if the enhancement is too strong, it can easily amplify noise, cause edge artifacts, or distort details; and their adaptability to complex non-uniform illumination and severely low-light scenes is also limited.

[0007] (4) Deep learning-based methods (CNN, GAN, etc.): CNN is better at extracting multi-level features, improving brightness, restoring details, and suppressing noise through end-to-end training; GAN, on the other hand, uses adversarial learning between the generator and the discriminator to make the enhancement results closer to natural images in terms of brightness, texture, and visual realism. These methods are highly adaptable to complex scenes and non-uniform lighting, and their overall performance is usually better than traditional methods. However, their limitations lie in their strong dependence on high-quality training data, high model training costs, weak interpretability, and the tendency to suffer from insufficient generalization, color distortion, artifacts, or unstable details when applied across different scenes.

[0008] In recent years, diffusion models have attracted widespread attention due to their excellent performance in tasks such as image generation, image inpainting, and detail reconstruction. Their core idea is to progressively add noise to the data and then learn an inverse denoising process to gradually recover a high-quality image from pure noise. Compared to traditional generative models, diffusion models have advantages such as relatively stable training processes, richer details in generated results, and stronger distribution fitting capabilities, especially excelling in texture generation, structural completion, and visual realism. Therefore, introducing diffusion models into low-light image enhancement is considered a promising research direction, expected to improve the ability to restore details in dark areas and overall visual quality.

[0009] However, directly applying diffusion models to low-light image enhancement still presents several challenges. First, the generation process typically relies on multi-step iterative denoising, resulting in high computational costs and slow inference speeds, making it difficult to meet the demands of real-time processing or edge device deployment. Second, low-light enhancement requires not only increasing brightness but also accurately preserving the structural and textural information of the original image. However, diffusion models, during repeated denoising and reconstruction, are prone to edge blurring, texture loss, and even detail forgery. Third, low-light images are often accompanied by complex noise, with significant differences in noise levels across different scenes. Traditional diffusion frameworks lack adaptive modeling capabilities for noise strength and type, easily leading to insufficient denoising or over-smoothing. Finally, low-light enhancement also involves balancing brightness enhancement and color restoration. If the model does not adequately model the coupling relationship between illumination and color, it may cause color cast, saturation imbalance, or overall color distortion while enhancing brightness. Therefore, reducing the inference cost of diffusion models, enhancing their structure preservation capabilities, and improving their adaptive control over noise and color have become key issues in current low-light image enhancement research. Summary of the Invention

[0010] The purpose of this application is to provide a dynamic residual hybrid expert low-light image enhancement method, apparatus, and storage medium.

[0011] To achieve the above objectives, the first aspect of this application provides a dynamic residual hybrid expert low-light image enhancement method, the method comprising:

[0012] The low-light image to be enhanced is subjected to feature decomposition to obtain potential illumination features for characterizing illumination information and potential reflection features for characterizing the inherent texture information of the image.

[0013] Based on the feature decomposition results, combined with the spatial confidence map, diffusion time step, illumination prior center and reflection prior center, a dynamic prior center field for the current diffusion time step is generated. The dynamic prior center field changes with the time step and spatial location. The illumination prior center and the reflection prior center are constructed from latent illumination features and latent reflection features, respectively.

[0014] The noisy features of the current time step, the diffusion time step, the spatial confidence map, the reflection prior center, and the dynamic prior center field are input into the improved denoising network, and residual prediction is performed based on the dynamic prior center field within the denoising network to obtain the predicted spatial residual.

[0015] By superimposing the dynamic prior central field with the predicted spatial residual, and then iteratively diffused, the clear latent features of the reflection component are obtained.

[0016] The enhanced latent features of the sharp reflection component and the enhanced latent features of the illumination component are fused to obtain an enhanced latent representation, and the enhanced latent representation is decoded to obtain the final enhanced image; wherein the enhanced latent features of the illumination component are obtained by enhancing the latent illumination features.

[0017] In this embodiment of the application, before the step of performing feature decomposition processing on the low-light image to be enhanced to obtain potential illumination features for characterizing illumination information and potential reflection features for characterizing the inherent texture information of the image, the method further includes:

[0018] The low-light image to be enhanced is input into the first convolution module, and the basic bright and dark edge features and local texture features of the image are extracted through convolution mapping to obtain the initial feature map;

[0019] The initial feature map is sequentially input into a multi-level downsampling module, which contains at least two downsampling units. Each downsampling unit consists of a convolutional layer or a pooling layer with a stride of 2, used to progressively compress the spatial resolution of the feature map.

[0020] At least one residual block is inserted before and after each downsampling operation. The residual block contains skip connections to compensate for the loss of structural information caused by downsampling during the reduction of spatial resolution and to maintain gradient flow stability.

[0021] After being processed by the multi-level downsampling module, the resulting feature map is input into a 1×1 convolutional layer. The feature channels are linearly combined and dimensionality reduced through pointwise convolution, and the feature map is mapped to a unified latent space dimension to obtain a low-dimensional latent representation of the low-light image to be enhanced.

[0022] In this embodiment of the application, the step of performing feature decomposition processing on the low-light image to be enhanced to obtain potential illumination features for characterizing illumination information and potential reflection features for characterizing the inherent texture information of the image includes:

[0023] The low-dimensional latent representation is input into a shared backbone network to perform deep feature mapping on the low-dimensional latent representation and extract shared backbone features that simultaneously contain structural information, texture information and lighting information.

[0024] The shared backbone features are input in parallel to the reflectivity branch network and the illumination branch network, respectively. The reflectivity branch network is used to decouple from the shared backbone features and output the latent reflectivity features, which are used to characterize the texture, edges, and inherent structure information of objects in the image. The illumination branch network is used to decouple from the shared backbone features and output the latent illumination features, which are used to characterize the non-uniform illumination distribution and brightness variation information in the scene.

[0025] In this embodiment of the application, the step of generating a dynamic prior center field for the current diffusion time step based on the feature decomposition result, combined with the spatial confidence map, diffusion time step, illumination prior center, and reflection prior center includes:

[0026] Spatial confidence maps are generated based on illumination prior centers;

[0027] Based on the spatial confidence map and combined with the diffusion time step, the illumination prior center and the reflection prior center are spatially fused by position-by-position adaptive weighting to construct the dynamic prior center field of the current diffusion time step.

[0028] In this embodiment of the application, after generating the dynamic prior central field of the current diffusion time step, the method further includes:

[0029] Based on the dynamic prior center field, the noisy latent features in the diffusion process are reparameterized so that the sampling center of the diffusion sampling process is shifted from the random noise center to the prior center containing basic image contour information and coarse structure information.

[0030] At different diffusion time steps and different spatial locations, the dynamic prior central field is adaptively adjusted according to the spatial confidence map and diffusion time step, so that the dynamic prior central field has different prior guidance intensities in dark light regions, structural regions and texture regions.

[0031] Using the reparameterized dynamic prior central field as the benchmark for spatial residual prediction, the improved denoising network predicts the residual offset of the current noisy latent feature relative to the dynamic prior central field, and recursively obtains the clear reflection component latent feature based on the superposition result of the dynamic prior central field and the residual offset.

[0032] In this embodiment of the application, the step of inputting the noisy features of the current time step, the diffusion time step, the spatial confidence map, the reflection prior center, and the dynamic prior center field into the improved denoising network, and performing residual prediction based on the dynamic prior center field within the denoising network to obtain the predicted spatial residual, includes:

[0033] The denoising network includes a local time modulation sparse routing module, which distributes the image information of the current diffusion time step to the heterogeneous expert network based on the spatial confidence information constructed according to the local illumination conditions.

[0034] The local time modulation sparse routing module combines the current noisy features, diffusion time steps, reflection prior centers, spatial confidence maps, and dynamic prior center field values ​​to generate the allocation probability distribution corresponding to each expert network in the heterogeneous expert network.

[0035] The k expert networks with the highest allocation probability are selected as the activated expert networks based on the sparse Top-k selection method.

[0036] The probabilities of the activated expert network are normalized to obtain the final gating weights.

[0037] The activated expert network processes the features of the received image to be processed and then performs a weighted summation using the corresponding gating weights to obtain the predicted spatial residual.

[0038] In this embodiment of the application, the step of the local time modulation sparse routing module generating the allocation probability distribution corresponding to each expert network in the heterogeneous expert network by combining the current noisy features, diffusion time steps, reflection prior centers, spatial confidence maps, and dynamic prior center field values ​​includes:

[0039] The noisy features of the current time step, the diffusion time step, the reflection prior center, the spatial confidence map, and the dynamic prior center field are fused to construct a comprehensive routing context vector;

[0040] The integrated routing context vector is input into a routing network composed of two layers of sensing mechanisms containing activation functions, and is mapped to an allocation probability distribution of N heterogeneous expert sub-networks through the Softmax function.

[0041] In this embodiment, the improved denoising network is an asymmetric multi-scale diffusion denoising network, which includes a network layer with a hybrid expert module and a network layer without a hybrid expert module.

[0042] The network layer with the hybrid expert module is set in the bottleneck layer and the medium-low resolution layer after downsampling, and is used to perform differential feature recovery on dark light areas, noisy areas or structurally missing areas.

[0043] The network layer without the hybrid expert module is set in the high-resolution layer and the shallow feature layer, and is used to perform basic denoising and detail preservation processing on bright areas and texture fidelity areas.

[0044] The network layer with the hybrid expert module and the network layer without the hybrid expert module are arranged asymmetrically in a multi-scale hierarchy from shallow to deep.

[0045] In this embodiment of the application, the method further includes:

[0046] The training process of the dynamic residual hybrid expert low-light image enhancement method includes applying constraint losses to multiple preset key computing nodes. Each constraint loss is calculated based on the difference between the output features and target features of the corresponding node, and the weighted sum of all constraint losses is used as the total loss function to jointly optimize the model parameters.

[0047] A second aspect of this application provides a dynamic residual mixing expert low-light image enhancement apparatus, comprising: a memory configured to store instructions; and a processor configured to retrieve instructions from the memory and, when executing the instructions, to implement the dynamic residual mixing expert low-light image enhancement method.

[0048] A third aspect of this application provides a machine-readable storage medium storing instructions for causing a machine to perform a dynamic residual mixing expert low-light image enhancement method.

[0049] The above technical solution breaks through the traditional diffusion model denoising network and proposes an asymmetric multi-scale diffusion denoising backbone network for low-light image enhancement. The denoising backbone simultaneously incorporates networks with and without hybrid expert modules, and gated sparse routes are strategically deployed in the downsampled bottleneck layer and low-to-medium resolution layers to reduce overall inference computation. During the backdiffusion recovery process, the denoising network does not directly predict the full representation of the target's sharp latent features. Instead, it outputs the spatial residual offset relative to the prior benchmark for the current position and constructs the sharp feature estimate at the current moment based on this residual offset, further completing the recursive update of the latent state from the previous time step. Through the combination of this asymmetric architecture deployment and the residual recursive recovery mechanism, dark areas can utilize heterogeneous experts to perform significant structural repair with lower computational overhead, while bright and textured areas undergo small-amplitude fidelity denoising. This achieves differentiated enhancement of different regions of the low-light image while balancing inference efficiency, recovery stability, and detail fidelity.

[0050] Other features and advantages of the embodiments of this application will be described in detail in the following detailed description section. Attached Figure Description

[0051] The accompanying drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the following detailed description to explain the embodiments of this application, but do not constitute a limitation on the embodiments of this application. In the drawings:

[0052] Figure 1 The schematic diagram illustrates a flow chart of a dynamic residual hybrid expert low-light image enhancement method according to an embodiment of this application;

[0053] Figure 2 A schematic diagram illustrating the dimensionality reduction and decoupling module block according to an embodiment of this application is shown.

[0054] Figure 3 The diagram illustrates an asymmetric multi-scale diffusion denoising network structure according to an embodiment of this application.

[0055] Figure 4A schematic diagram of the LU-Moe Block structure according to an embodiment of this application is shown.

[0056] Figure 5 This illustration schematically depicts a routing mechanism and a hybrid expert model according to embodiments of this application;

[0057] Figure 6 The network structure of the image restoration module according to an embodiment of this application is illustrated schematically;

[0058] Figure 7 The illustration shows a comparison of low-light image enhancement results obtained according to embodiments of this application;

[0059] Figure 8 A schematic diagram of a computer device according to an embodiment of this application is shown. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only for illustration and explanation of the embodiments of this application and are not intended to limit the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0061] Figure 1 The illustration schematically shows a flowchart of a dynamic residual hybrid expert low-light image enhancement method according to an embodiment of this application. Figure 1 As shown in the embodiments of this application, a dynamic residual hybrid expert low-light image enhancement method is provided, which may include the following steps:

[0062] Step 10: Perform feature decomposition on the low-light image to be enhanced to obtain potential illumination features for characterizing illumination information and potential reflection features for characterizing the inherent texture information of the image.

[0063] Before enhancing the low-light image, the original image is first subjected to dimensionality reduction. The purpose of dimensionality reduction is twofold: firstly, to suppress sensor noise and artifacts generated under extremely low lighting conditions, preventing these interferences from being amplified in subsequent decomposition and enhancement processes; and secondly, to reduce the spatial resolution or channel dimension of image features, thereby reducing the computational overhead of subsequent illumination-reflectivity decomposition and diffusion models, and improving overall processing efficiency. Specifically, learnable convolutional projection layers can be used to map the input image from its original high-dimensional space to a low-dimensional, compact representation. The dimensionality-reduced features retain the main structural information and brightness distribution of the scene while filtering out high-frequency noise and redundant details, providing a cleaner input for accurate illumination and reflectivity decomposition. Dimensionality reduction is a preparatory step for the diffusion process; reducing the dimensionality of the low-light image reduces computational load, while preserving as much of the scene's main structural, edge, and semantic information as possible during compression.

[0064] In this embodiment, see Figure 2 This is a block diagram of the dimensionality reduction and decoupling module in this embodiment. Specific dimensionality reduction methods include:

[0065] Step S1011: Input the low-light image to be enhanced into the first convolution module, and perform channel-by-channel convolution operation on the input image through the convolution kernel to extract the basic bright and dark edge features and local texture features of the image to obtain the initial feature map;

[0066] Step S1012: The initial feature map is sequentially input into a multi-level downsampling module. The multi-level downsampling module contains at least two downsampling units. Each downsampling unit consists of a convolutional layer or pooling layer with a stride of 2, used to progressively compress the spatial resolution of the feature map. At least one residual block is inserted before and after each downsampling operation. The residual block contains skip connections to compensate for the loss of structural information caused by downsampling during the reduction of spatial resolution and to maintain gradient flow stability.

[0067] Step S1013: After processing by the multi-level downsampling module, the obtained feature map is input into a 1×1 convolutional layer. The feature channels are linearly combined and dimensionality reduced by point-by-point convolution, and the feature map is mapped to a unified latent space dimension to obtain a low-dimensional latent representation of the low-light image to be enhanced. The low-dimensional latent representation is used as the prior input for subsequent illumination-reflectivity decomposition or diffusion model.

[0068] Specifically, the input to the dimensionality reduction module is a low-light image. Low-light images The input is fed into a dimensionality reduction network to obtain a low-dimensional latent representation. Basic bright and dark edges and local texture features are extracted through convolutional mapping. Then, the spatial resolution is progressively compressed through multi-level downsampling, with residual blocks inserted before and after each downsampling stage to reduce structural information loss during compression. Finally, 1*1 convolutions are used to map the features to a unified latent dimension to obtain the low-dimensional latent representation. The formula is as follows:

[0069]

[0070] in, This represents the input low-light image. (.) indicates a dimension reduction encoder. This represents the low-dimensional latent representation output by the encoder. This formula is used to characterize the mapping process of a low-light image from the original pixel space to the latent feature space. Its purpose is to preserve as much of the main structural, edge, and semantic information in the scene as possible while compressing the spatial resolution and reducing the computational cost of subsequent diffusion restoration.

[0071] Furthermore, in this embodiment, the reduced image features are further decomposed to separate and model the reflection information and illumination information that were originally coupled in the latent representation of the low-light image, so that the subsequent diffusion network can apply conditional constraints for structure restoration and brightness correction respectively.

[0072] The specific decomposition method includes: inputting the low-dimensional latent representation into a shared backbone network, wherein the shared backbone network is used to perform deep feature mapping on the low-dimensional latent representation and extract shared backbone features that simultaneously contain structural information, texture information and lighting information;

[0073] The shared backbone features are input in parallel to the reflectivity branch network and the illumination branch network, respectively. The reflectivity branch network is used to decouple from the shared backbone features and output latent reflectivity features, which are used to characterize the inherent texture information and detail structure of the image. The illumination branch network is used to decouple from the shared backbone features and output latent illumination features, which are used to characterize the non-uniform illumination distribution in the scene.

[0074] Specifically, see Figure 2 The low-dimensional latent representation obtained after processing in the dimensionality reduction module The input is fed into the decoupling module to convert the low-dimensional latent representation. The shared backbone network is input, and unified basic features are extracted from the mixed structure, texture, and lighting information present within it to obtain the shared basic features, i.e., the shared backbone features. These provide a common representation foundation for the subsequent two specialized branches. The formula is as follows:

[0075]

[0076] in, Indicates sharing the backbone network. This represents the shared backbone features. The shared backbone network first represents the low-dimensional latent features. A unified basic feature extraction is performed to obtain a common representation that simultaneously contains structural, textural, and lighting information, providing a common input basis for the reflectivity and illumination branches.

[0077] Latent space decoupling module: The shared backbone features are input into the reflectivity branch network R-Branch and the illumination branch network L-Branch, respectively. The illumination branch L-Branch outputs the latent illumination features Z. L The reflectivity branch network R-Branch outputs the latent reflectivity feature Z. R .

[0078] This dual-branch decoupling method separates and models the reflection and illumination information that are coupled in the latent representation of low-light images. This allows the subsequent diffusion network to apply conditional constraints for structure restoration and brightness correction, respectively. The formula is as follows:

[0079]

[0080] in, , These are the reflectivity branch network R-Branch and the illumination branch network L-Branch, respectively. Latent reflectance feature Z R Z is used to characterize the structure, texture, and inherent properties of objects in an image; it is a latent illumination feature. L Used to characterize the distribution and brightness changes of lighting in a scene.

[0081] Step 20: Based on the feature decomposition results, and combining the spatial confidence map, diffusion time step, illumination prior center, and reflection prior center, generate a dynamic prior center field for the current diffusion time step. The dynamic prior center field changes with the time step and spatial location. The illumination prior center and the reflection prior center are constructed from latent illumination features and latent reflection features, respectively.

[0082] The specific steps include:

[0083] Step 1021: Generate a spatial confidence map based on the illumination prior center;

[0084] Step 1022: Based on the spatial confidence map and the diffusion time step, adaptive weighted fusion of the illumination prior center and the reflection prior center is performed at the spatial location (x,y) to construct a dynamic prior center field.

[0085] By shifting the distribution of the diffusion process, the starting point of the denoising process is translated to a physical prior center that already has the basic image contour. It should be noted that the physical prior center here and the "dynamic prior center" mentioned later refer to the same technical object. The use of "physical prior center" here indicates that the starting point of the denoising process is no longer pure random noise, but rather a prior center that has been shifted to a basic image contour. Later, it is further explained that this prior center is based on a spatial confidence map and is dynamically changing in spatial location over time. Therefore, using the same identifier in both instances is to express different levels of description of the same prior center quantity, rather than representing two different contents.

[0086] The denoising method in this application denoises based on an existing foundation, rather than directly denoising an original image. The aforementioned physical prior center means that the denoising network does not predict the full amount of information, but instead constructs a dynamic prior center field that evolves with space and time based on the spatial confidence map. Combined with the gating routing mechanism in the multi-scale hierarchical denoising network architecture, it predicts the heterogeneous spatial residuals pointing from this dynamic center to the real clear features, and finally reconstructs high-quality enhanced latent features step by step through the sampling sequence.

[0087] The physical priori center here It is obtained by position-wise adaptive fusion of the illumination prior center and the reflection prior center under the constraints of spatial confidence map and diffusion time step. Essentially, it is a prior guiding center containing basic image contour and coarse structure information, which is used to shift the denoising starting point from pure random noise to the sampling center with basic structure prior.

[0088] To address the problem of error accumulation caused by the destruction of reflectivity characteristics in extremely low illumination regions,

[0089] (1) First, using the prior center of illumination Z L Generate spatial confidence map :

[0090]

[0091] in, It is the sigmoid activation function. For a lightweight mapping network used to aggregate channel information, . This indicates the degree of dependence or confidence of the position (x,y) on the reflection prior center at diffusion step t. The larger the value, the more likely the position is to adopt the reflection prior center; the smaller the value, the more likely the position is to adopt the illumination prior center.

[0092] (2) Subsequently, combined with the diffusion time step t, the reflection prior center Z at the spatial location (x,y) is... RAnd the Lighting A priori Center Z L Adaptive weighted fusion is performed to construct a dynamic prior central field. :

[0093]

[0094] in, This represents the Hadamard product (element-by-element multiplication). and The time annealing parameter is used to control the proportion of physical prior intervention. This formula is used to generate a dynamic prior center field: at each diffusion step and at each spatial location, an adaptive weighted fusion is performed between the reflection prior center and the illumination prior center based on the spatial confidence map, and the participation degree of the two types of priors at different diffusion stages is controlled by the time annealing parameter, so that the dark area is biased towards illumination correction and the structure area is biased towards reflectivity recovery.

[0095] That is to say, after generating the dynamic prior central field of the current diffusion time step, the process also includes:

[0096] Based on the dynamic prior center field, the noisy latent features in the diffusion process are reparameterized so that the sampling center of the diffusion sampling process is shifted from the random noise center to the prior center containing basic image contour information and coarse structure information.

[0097] At different diffusion time steps and different spatial locations, the dynamic prior central field is adaptively adjusted according to the spatial confidence map and diffusion time step, so that the dynamic prior central field has different prior guidance intensities in dark light regions, structural regions and texture regions.

[0098] Using the reparameterized dynamic prior central field as the benchmark for spatial residual prediction, the improved denoising network predicts the residual offset of the current noisy latent feature relative to the dynamic prior central field, and recursively obtains the clear reflection component latent feature based on the superposition result of the dynamic prior central field and the residual offset.

[0099] In summary, this application innovatively proposes constructing a priori reflection centers and a priori illumination centers based on latent reflection and latent illumination features. It then utilizes a spatial confidence map to adaptively and weightedly fuse these two centers position-by-position, generating a dynamic priori center field corresponding to the current diffusion step. This dynamic priori center field is used to reparameterize diffusion variables, enabling the diffusion sampling center to dynamically adjust with spatial location and diffusion stage. This results in a greater bias towards illumination correction in underexposed dark areas and towards reflectivity recovery in structural regions, improving the recovery stability and priori adaptation capability in complex low-light scenes. By achieving dynamic adjustment for different regions, it effectively enhances the adaptive adjustment capability of the diffusion model in noisy, dark areas.

[0100] Step 30: Input the noisy features of the current time step, the diffusion time step, the spatial confidence map, the reflection prior center, and the dynamic prior center field into the improved denoising network. After processing by the denoising network, the predicted spatial residual is obtained. That is, input the noisy features of the current time step, the diffusion time step, the spatial confidence map, the reflection prior center, and the dynamic prior center field into the improved denoising network, and perform residual prediction based on the dynamic prior center field within the denoising network to obtain the predicted spatial residual.

[0101] Noisy features at the current time step Diffusion time step t, spatial confidence plot Z, the prior center of reflection R and dynamic prior central field Input to a denoising network with a gated sparse network In the middle, the spatial residual of the output prediction :

[0102]

[0103] In the formula: This represents the noisy feature at the current time step, where t is the time step of the current diffusion process. For spatial confidence maps, For noise reduction networks, These are the prior centers of reflection, This is a complete spatial residual map. In this embodiment, the reflection prior center is directly constructed from the latent reflection features, so the same notation is used to represent it. It should be noted that the same notation is used in this embodiment because "latent illumination features / latent reflection features" and "illumination prior center / reflection prior center" are not two independent sets of variables under this scheme, but rather different functional descriptions of the same representation at different processing stages. Specifically, the latent illumination features and latent reflection features output by the decoupling module are directly used as the corresponding illumination prior center and reflection prior center, respectively, in the subsequent construction of the dynamic prior center field, without introducing new mapping variables or additional center parameters. Among them, the reflection prior center serves as the external condition input of the denoising network, while the dynamic prior center field serves as the prior benchmark for residual prediction within the denoising network. Therefore, to maintain the simplicity of the notation and highlight the continuity of the technical process, this paper uses corresponding notation to represent the relevant representations. That is, in the context of feature decoupling, it represents the latent feature; in the context of prior construction, it represents the prior center directly undertaken by the latent feature. In this embodiment, the two have a one-to-one correspondence, so no new symbols are introduced. The formula uses conditional inputs containing reflection features and temporal information to predict the residual shift of the target's sharp potential features relative to the current state. Therefore, the diffusion recovery process in low-light image enhancement can be modeled as a conditionally constrained residual recursive process, enabling the network to adaptively perform differentiated recovery in different illumination regions, which is particularly beneficial for structural reconstruction and texture fidelity in extremely dark areas.

[0104] At this point, the estimated value of the noise-free sharp features (i.e., the target reflectivity manifold) at the current moment. Defined as the spatial superposition of the dynamic prior central field and the residuals of the network prediction space: .

[0105] Step 40: The dynamic prior central field is superimposed with the predicted spatial residual, and the sharp reflection component latent features are obtained after iterative diffusion. During the training phase, the noisy features of the current diffusion time step are obtained by adding noise to the sharp reflection component latent features corresponding to the target normal light image according to a preset noise schedule. During the inference phase, a dynamic prior central field is constructed based on the latent reflection features and latent illumination features obtained by decoupling the low-light image to be enhanced, and the sharp reflection component latent features are obtained by back-diffusion recursively using this dynamic prior central field as the residual prediction benchmark.

[0106] Specifically, the estimate of the noise-free sharp features (i.e., the target reflectivity manifold) at the current moment. Defined as the spatial superposition of the dynamic prior central field and the network prediction residuals: .

[0107] Using the obtained clear feature estimates The latent feature state of the previous time step is calculated by combining the diffusion posterior transition probability. :

[0108]

[0109] in, The noise scheduling cumulative variance parameter is preset for the diffusion process. This is the superposition of the dynamic prior central field and the network prediction residuals. Let be the latent characteristic state at time t. Let t be the latent feature state at time t-1. By iteratively performing the above state update, the reflection component latent feature with clear and realistic texture is finally output at time t=0. .

[0110] Step 50: The enhanced sharp reflection component latent features and the enhanced illumination component latent features are fused to obtain an enhanced latent representation, and the enhanced latent representation is decoded to obtain the final enhanced image; wherein the enhanced illumination component latent features are obtained by enhancing the latent illumination features.

[0111] See Figure 6 The image restoration module is used to extract the enhanced latent features of the reflection component obtained from diffusion restoration. With enhanced latent characteristics of illumination components The images are fused to construct a complete enhanced latent representation, and then mapped back to the image space using an image decoder to obtain the final enhanced image. .

[0112] The decoupled ultra-dark potential illumination feature Z L For brightening, a dedicated lightweight brightening mapping network is used. The network mainly consists of multiple stacked depthwise separable convolutional layers and nonlinear activation functions; when illumination features are input into the network, the enhanced latent features of the illumination components are directly output. Denoising the latent features of the reflection component Latent features of illumination components after convolution brightening By performing splicing and feature alignment, the fused low-dimensional enhanced latent features are obtained. ,Will The input is fed into a decoder network consisting of multiple layers of upsampling operators and residual convolutional blocks. In this process, it is mapped back to the high-resolution RGB pixel space:

[0113]

[0114] in, This is the final output low-light enhanced image. This is a low-dimensional enhancement feature.

[0115] This invention provides a dynamic residual hybrid expert low-light image enhancement method, comprising: performing feature decomposition on the low-light image to be enhanced to obtain latent illumination features for characterizing illumination information and latent reflection features for characterizing inherent texture information of the image; based on the feature decomposition results, combining a spatial confidence map, a diffusion time step, an illumination prior center, and a reflection prior center, generating a dynamic prior center field for the current diffusion time step, wherein the dynamic prior center field changes with the time step and spatial position, and the illumination prior center and the reflection prior center are constructed from latent illumination features and latent reflection features, respectively; inputting the noisy features of the current time step, the diffusion time step, the spatial confidence map, the reflection prior center, and the dynamic prior center field into an improved denoising network, and performing residual prediction based on the dynamic prior center field within the denoising network to obtain the predicted spatial residual; superimposing the dynamic prior center field and the predicted spatial residual, and obtaining a clear reflection component latent feature after iterative diffusion; fusing the reflection component latent feature with the illumination component latent feature to reconstruct the final enhanced image features. The method in this application addresses the common problems in low-light images, such as insufficient brightness, amplified noise, local underexposure, color distortion, and loss of texture details. It constructs reflection-related and illumination-related representations in the latent space and combines an illumination-guided dynamic prior reparameterization mechanism with a spatiotemporal routing hybrid expert denoising backbone network to achieve adaptive enhancement and restoration of low-light images.

[0116] This application provides a low-light image enhancement method to address the technical problems in existing low-light image enhancement techniques, such as detail loss, color distortion, and the tendency of traditional diffusion models to fail in extremely low-light regions due to their fixed sampling starting point. This application selects the publicly available low-light image dataset LOL-v1 and compares its PSNR and SSIM metrics with other methods. The specific results are shown in Table 1. ,

[0117] This application outperforms the aforementioned comparative methods in all PSNR metrics and ranks second in the SSIM (Structural Similarity Index), effectively demonstrating the excellent effect of this design in low-light image enhancement. Specific enhanced image results are shown below. Figure 7 As shown.

[0118] In one embodiment, refer to Figures 3-5 It further describes how the improved denoising network enhances images.

[0119] Specifically, the denoising network with gated sparse network in this embodiment This is an improved denoising network that employs a routing mechanism and a hybrid expert model.

[0120] The step of inputting the noisy features of the current time step, the diffusion time step, the spatial confidence map, the reflection prior center, and the dynamic prior center field into the improved denoising network, and obtaining the predicted spatial residual after processing by the denoising network, includes the following steps:

[0121] Step 1031: The denoising network includes a local time modulation sparse routing module, which distributes features to a heterogeneous expert network based on local illumination conditions.

[0122] Step 1032: The local time modulation sparse routing module combines the current noisy features, diffusion time steps, reflection prior centers, spatial confidence maps, and dynamic prior center field values ​​to generate the allocation probability distributions corresponding to each expert network in the heterogeneous expert network. Specifically, it combines the current noisy features, diffusion time steps, reflection prior centers, spatial confidence maps, and dynamic prior center field values ​​to construct a comprehensive routing context vector. The comprehensive routing context vector is then input into a routing network composed of two layers of perceptrons containing activation functions, and mapped to the allocation probability distributions of N heterogeneous expert sub-networks using the Softmax function.

[0123] Step 1033: Select the k expert networks with the highest allocation probability according to the sparse Top-k selection method and determine them as the activated expert networks;

[0124] Step 1034: Normalize the probabilities of the activated expert network to obtain the final gating weights.

[0125] Step 1035: The activated expert network processes the features of the received image to be processed and obtains the predicted spatial residual by weighted summation using the corresponding gating weights.

[0126] The above steps are described in detail as follows: In this embodiment, an improved denoising network... In the process, the routing mechanism distributes the image information of the current diffusion time step to heterogeneous experts based on the reflection prior center, spatial confidence map, dynamic prior center field value, and diffusion time step, referring to... Figure 3 and Figure 4 To reduce the computational cost of the model, this application uses two different backbone network structures in the diffusion model denoising network: a U-net-like structure network with an expert model and one without. The approximate model of the LU-moe Block is as follows: Figure 4As shown, the LU Block simply replaces the hybrid expert layers within the model with FFN layers. The network layers with hybrid expert modules are placed in the downsampled bottleneck layer and low-to-medium resolution layer, used to perform differential feature recovery in dark, noisy, or structurally deficient regions. The network layers without hybrid expert modules are placed in the high-resolution layer and shallow feature layer, used to perform basic denoising and detail preservation processing in bright and texture-preserving regions. This asymmetric combination of network layers with and without hybrid expert modules reduces the overall inference computation and improves the differential recovery capability of different regions in low-light images.

[0127] The local temporal modulation sparse routing module, for location units (tokens) at different spatial locations, combines the location unit features (token features) of the current location, the diffusion time step, the reflection prior center, the spatial confidence map, and the dynamic prior center field value to generate routing scores for each expert network. Here, the token feature refers to the high-dimensional feature representation of a local location in the image. In this embodiment, the routing mechanism "determines which expert network to assign to a token at a different spatial location by combining the token features of the current location, the time step, the reflectivity guidance features, the spatial confidence map, and the dynamic prior center field value." After an image is patched, it is no longer directly viewed as the original pixels, but rather becomes a feature representation of many "small blocks" or "location points." Each location corresponds to a vector, which is called the token feature of that location. After generating routing scores for each expert network, the activated expert network and its corresponding weights are determined using a sparse Top-k selection method. Then, the outputs of the activated expert networks are weighted and summed according to the weights to achieve differentiated recovery for different diffusion stages and different spatial regions. Within this residual translation framework, dark area features are routed to a structure generation expert to predict large-scale structure repair residuals, while bright area features are routed to a high-frequency denoising expert to predict small-scale texture fidelity residuals.

[0128] Reference Figures 3-5 The above process is described in further detail below:

[0129] Current noisy features diffusion time step Reflection Prior Center Spatial confidence map and dynamic prior central field value Concatenate the data along the channel dimension to construct a comprehensive routing context vector. That is, the integrated routing context vector. Packed the current noisy features diffusion time step Reflection Prior Center Spatial confidence map and dynamic prior central field value This information is combined into a long vector. .

[0130] Then, the context vector The input is fed into a routing network consisting of a two-layer perceptron (MLP) containing activation functions, and then mapped to the original probability distribution of N heterogeneous expert subnetworks through the Softmax function. :

[0131]

[0132] In the formula, and The learnable weight matrix for the routing network. ReLU is a non-linear activation function. This represents the probability that the i-th feature block is assigned.

[0133] To reduce computational overhead and ensure network sparsity, this embodiment employs TOP-k expert subnetworks. The process involves forward processing. The generated raw allocation probabilities are sorted, and only the k experts with the highest probability values ​​(i.e., the set of activated experts) are retained. The probabilities of the remaining k experts are forcibly truncated to 0. Then, the probabilities of the retained k experts are renormalized to obtain the final gating weights. :

[0134]

[0135] In the formula, Let represent the probability of the i-th position relative to the j-th expert. This represents the sum of probabilities of the activated experts. Each expert network performs differentiated feature extraction and recovery based on local illumination characteristics. Finally, the output features of each activated expert are used with their corresponding gating weights. Perform weighted summation and output the aggregated residual features. :

[0136]

[0137] In the formula, Let be the gating weight corresponding to the i-th position. The enhanced feature output by the i-th position input to the j-th expert subnetwork.

[0138] The residual characteristics calculated at each spatial location Arranged and combined according to the original space, a complete spatial residual map is formed, as mentioned above. , which serves as the output of the denoising network.

[0139] This embodiment proposes a local temporal modulation sparse routing driven by reflection prior center features, spatial confidence maps, and dynamic prior center field values. When allocating experts, for each image patch in the feature sequence, its own feature vector, time step, spatial confidence map and dynamic prior center field value strictly aligned with its spatial coordinates are combined to output the probability distribution of the current image patch for each expert sub-network, thereby mobilizing the expert network to perform specific enhancements on the image. This mechanism ensures that high-frequency and low-frequency features of different illumination intensity regions within the same image can be distributed in parallel to heterogeneous expert sub-networks, realizing dynamic enhancement for different illumination intensity regions.

[0140] Overall, this application proposes an asymmetric multi-scale diffusion denoising backbone network for low-light image enhancement. The denoising backbone simultaneously incorporates networks with and without hybrid expert modules, and gated sparse routing is strategically deployed in the downsampled bottleneck layer and low-to-medium resolution layers to reduce overall inference computation. During backdiffusion recovery, the denoising network does not directly predict the full representation of the sharp latent features of the target. Instead, it outputs a spatial residual offset relative to the prior benchmark for the current position and constructs a sharp feature estimate for the current time step based on this residual offset, further recursively updating the latent state from the previous time step. Through the combination of this asymmetric architecture deployment and the residual recursive recovery mechanism, dark areas can utilize heterogeneous experts for significant structural restoration with lower computational overhead, while bright and textured areas undergo small-amplitude fidelity denoising. This achieves differentiated enhancement of different regions of the low-light image while balancing inference efficiency, recovery stability, and detail fidelity.

[0141] Furthermore, in another embodiment, this application further applies loss constraints to each stage of the entire model architecture. The training process of the dynamic residual hybrid expert low-light image enhancement method includes applying constraint losses to multiple preset key computation nodes. Each constraint loss is calculated based on the difference between the output features and target features of the corresponding node, and the weighted sum of all constraint losses is used as the total loss function to jointly optimize the model parameters. Specifically, the constraints include the following:

[0142] During the training phase, the following loss function is applied to the above decoupling modules to constrain the physical correctness of the decoupling:

[0143] 1. Dimensionality reduction and decoupling module losses ( )

[0144] In the latent space, the feature is decoupled into the reflectivity component Z. R and illumination component Z LTo ensure the correctness of decoupling, it is necessary to introduce physical prior constraints:

[0145] (1) Loss of reflectivity uniformity Extract the normal illumination image corresponding to the training process, and the extracted reflectance features are: The reflectivity uniformity loss is then: .

[0146] In the formula, As a potential reflection feature, This represents the reflectance characteristics of a real image.

[0147] (2) Loss of smoothness of illumination Real-world illumination distributions are typically locally smooth. Using a gradient-based smoothness penalty and leveraging the gradient of reflectivity as weights, the illumination smoothness loss is: .

[0148] In the formula: The spatial coordinates of the image feature map in the horizontal and vertical directions. These represent the gradients (i.e., the differences between adjacent pixels) in the horizontal and vertical directions, respectively. For potential reflection characteristics and potential illumination characteristics, For exponentially decaying weight terms, It is a hyperparameter that controls the smoothness sensitivity. This loss is calculated by adjusting Z... L We use a weighted summation of the absolute values ​​of the gradients to constrain the illumination components to change as slowly as possible in space, thereby suppressing unreasonable high-frequency fluctuations and local noise.

[0149] The total loss of the decoupling module is: .

[0150] In the formula: The total loss of the decoupling module, For the loss of consistency of reflection components, To address the loss of illumination smoothness, a consistency constraint is imposed on the reflection component and a smoothness constraint is imposed on the illumination component. The network is then jointly guided to learn a physically interpretable potential decoupled representation to improve the accuracy of structure preservation and illumination correction during low-light image enhancement.

[0151] 2. Dynamic residual diffusion denoising loss ( )

[0152] Residual estimation loss: True and clear latent reflection features are Then, at time step t, the target true residual should be The training objective of the diffusion denoising network is to minimize the network's prediction residuals. Mean square error between the true residual and the actual residual:

[0153]

[0154] In the formula: For the prediction residuals of the denoising network, This represents the true residual of the reflection component. This represents the dynamic residual diffusion denoising loss. This represents the expectation operation for the training samples, diffusion time steps, and random noise perturbations. This loss function drives the diffusion denoising network to learn a residual offset mapping oriented towards dynamic prior constraints by minimizing the squared error between the predicted and true residuals, thus transforming the backdiffusion recovery process in low-light image enhancement into a conditionally guided residual recursive optimization process.

[0155] 3. Load balancing losses in Hybrid Expert Network (MoE)

[0156] When training the MoE architecture, the phenomenon of "routing collapse" is highly likely to occur (i.e., all tokens flow to only a few expert networks, causing other experts to be idle). To ensure the effective utilization of sparse experts, load balancing loss must be introduced:

[0157] Where N represents the number of experts. This represents the proportion allocated to the i-th expert. Let be the average routing probability of the router for the i-th expert.

[0158] 4. Image Restoration and End-to-End Reconstruction Loss

[0159] The final enhanced image output by the decoder Need to match the target image Alignment at the pixel and perceptual levels results in the following reconstruction loss:

[0160]

[0161] in, This represents the image ultimately generated by the decoder. 1 represents the reference image (Ground Truth); 1 represents the L1 loss function.

[0162] Total loss function :

[0163] in, To balance the hyperparameters of the task weights, , , , Here are the loss functions for each module.

[0164] In one embodiment, a dynamic residual fusion expert low-light image enhancement device is provided. The dynamic residual fusion expert low-light image enhancement device includes a processor and a memory. The processor executes the program modules stored in the memory to achieve the corresponding functions.

[0165] The processor contains a kernel, which retrieves the corresponding program unit from memory. One or more kernels can be configured, and the dynamic residual blending expert low-light image enhancement method can be implemented by adjusting kernel parameters.

[0166] The memory may include non-permanent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0167] This application provides a storage medium storing a program that, when executed by a processor, implements the aforementioned dynamic residual mixing expert low-light image enhancement method.

[0168] This application provides a processor for running a program, wherein the program executes the above-described dynamic residual mixing expert low-light image enhancement method during runtime.

[0169] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 8 As shown. The computer device includes a processor A01, a network interface A02, memory (not shown), and a database (not shown) connected via a system bus. The processor A01 provides computing and control capabilities. The memory includes internal memory A03 and a non-volatile storage medium A04. The non-volatile storage medium A04 stores an operating system B01, a computer program B02, and a database (not shown). The internal memory A03 provides an environment for the operation of the operating system B01 and the computer program B02 stored in the non-volatile storage medium A04. The database is used to store data. The network interface A02 is used for communication with external terminals via a network connection. When executed by the processor A01, the computer program B02 implements a dynamic residual mixing expert low-light image enhancement method.

[0170] Those skilled in the art will understand that Figure 8The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0171] This application provides a computer (electronic) device, which includes a processor, a memory, and a program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of any of the above dynamic residual mixing expert low-light image enhancement methods.

[0172] This application also provides a computer program product, which, when executed on a data processing device, is used to perform the steps of the dynamic residual mixing expert low-light image enhancement method.

[0173] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0174] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0175] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0176] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0177] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0178] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0179] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0180] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0181] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A dynamic residual hybrid expert low-light image enhancement method, characterized in that, The method includes: The low-light image to be enhanced is subjected to feature decomposition to obtain potential illumination features for characterizing illumination information and potential reflection features for characterizing the inherent texture information of the image. Based on the feature decomposition results, combined with the spatial confidence map, diffusion time step, illumination prior center and reflection prior center, a dynamic prior center field for the current diffusion time step is generated. The dynamic prior center field changes with the time step and spatial location. The illumination prior center and the reflection prior center are constructed from latent illumination features and latent reflection features, respectively. The noisy features of the current time step, the diffusion time step, the spatial confidence map, the reflection prior center, and the dynamic prior center field are input into the improved denoising network. The dynamic prior center field is used as the residual prediction benchmark. After processing by the improved denoising network, the predicted spatial residual is obtained. By superimposing the dynamic prior central field with the predicted spatial residual, and then iteratively diffused, the clear latent features of the reflection component are obtained. The enhanced latent features of the sharp reflection component and the enhanced latent features of the illumination component are fused to obtain an enhanced latent representation, and the enhanced latent representation is decoded to obtain the final enhanced image; wherein the enhanced latent features of the illumination component are obtained by enhancing the latent illumination features.

2. The dynamic residual hybrid expert low-light image enhancement method according to claim 1, characterized in that, Before the step of performing feature decomposition processing on the low-light image to be enhanced to obtain potential illumination features for characterizing illumination information and potential reflection features for characterizing the inherent texture information of the image, the method further includes: The low-light image to be enhanced is input into the first convolution module, and the basic bright and dark edge features and local texture features of the image are extracted through convolution mapping to obtain the initial feature map; The initial feature map is sequentially input into a multi-level downsampling module, which contains at least two downsampling units. Each downsampling unit consists of a convolutional layer or a pooling layer with a stride of 2, used to progressively compress the spatial resolution of the feature map. At least one residual block is inserted before and after each downsampling operation. The residual block contains skip connections to compensate for the loss of structural information caused by downsampling during the reduction of spatial resolution and to maintain gradient flow stability. After being processed by the multi-level downsampling module, the resulting feature map is input into a 1×1 convolutional layer. The feature channels are linearly combined and dimensionality reduced through pointwise convolution, and the feature map is mapped to a unified latent space dimension to obtain a low-dimensional latent representation of the low-light image to be enhanced.

3. The dynamic residual hybrid expert low-light image enhancement method according to claim 2, characterized in that, The step of performing feature decomposition on the low-light image to be enhanced to obtain latent illumination features for characterizing illumination information and latent reflection features for characterizing the inherent texture information of the image includes: The low-dimensional latent representation is input into a shared backbone network to perform deep feature mapping on the low-dimensional latent representation and extract shared backbone features that simultaneously contain structural information, texture information and lighting information. The shared backbone features are input in parallel to the reflectivity branch network and the illumination branch network, respectively. The reflectivity branch network is used to decouple from the shared backbone features and output the latent reflectivity features, which are used to characterize the texture, edges, and inherent structure information of objects in the image. The illumination branch network is used to decouple from the shared backbone features and output the latent illumination features, which are used to characterize the non-uniform illumination distribution and brightness variation information in the scene.

4. The dynamic residual hybrid expert low-light image enhancement method according to claim 1, characterized in that, The step of generating the dynamic prior center field for the current diffusion time step based on the feature decomposition results, combined with the spatial confidence map, diffusion time step, illumination prior center, and reflection prior center, includes: Spatial confidence maps are generated based on illumination prior centers; Based on the spatial confidence map and combined with the diffusion time step, the illumination prior center and the reflection prior center are spatially fused by position-by-position adaptive weighting to construct the dynamic prior center field of the current diffusion time step.

5. The dynamic residual hybrid expert low-light image enhancement method according to claim 1 or 4, characterized in that, After generating the dynamic prior central field for the current diffusion time step, the process also includes: Based on the dynamic prior center field, the noisy latent features in the diffusion process are reparameterized so that the sampling center of the diffusion sampling process is shifted from the random noise center to the prior center containing basic image contour information and coarse structure information. At different diffusion time steps and different spatial locations, the dynamic prior central field is adaptively adjusted according to the spatial confidence map and diffusion time step, so that the dynamic prior central field has different prior guidance intensities in dark light regions, structural regions and texture regions. Using the reparameterized dynamic prior central field as the benchmark for spatial residual prediction, the improved denoising network predicts the residual offset of the current noisy latent feature relative to the dynamic prior central field, and recursively obtains the clear reflection component latent feature based on the superposition result of the dynamic prior central field and the residual offset.

6. The dynamic residual hybrid expert low-light image enhancement method according to claim 1, characterized in that, The steps of inputting the noisy features of the current time step, the diffusion time step, the spatial confidence map, the reflection prior center, and the dynamic prior center field into the improved denoising network, and using the dynamic prior center field as the residual prediction benchmark, to obtain the predicted spatial residual after processing by the improved denoising network include: The denoising network includes a local time modulation sparse routing module, which distributes the image information of the current diffusion time step to the heterogeneous expert network based on the spatial confidence map constructed according to the local illumination conditions. The local time modulation sparse routing module combines the current noisy features, diffusion time steps, reflection prior centers, spatial confidence maps, and dynamic prior center field values ​​to generate the allocation probability distribution corresponding to each expert network in the heterogeneous expert network. The k expert networks with the highest allocation probability are selected as the activated expert networks based on the sparse Top-k selection method. The probabilities of the activated expert network are normalized to obtain the final gating weights. The activated expert network processes the features of the received image to be processed and then performs a weighted summation using the corresponding gating weights to obtain the predicted spatial residual.

7. The dynamic residual hybrid expert low-light image enhancement method according to claim 6, characterized in that, The steps of the local time modulation sparse routing module in generating the allocation probability distribution for each expert network in the heterogeneous expert network by combining the current noisy features, diffusion time steps, reflection prior centers, spatial confidence maps, and dynamic prior center field values ​​include: The noisy features of the current time step, the diffusion time step, the reflection prior center, the spatial confidence map, and the dynamic prior center field are fused to construct a comprehensive routing context vector; The integrated routing context vector is input into a routing network composed of two layers of sensing mechanisms containing activation functions, and is mapped to an allocation probability distribution of N heterogeneous expert sub-networks through the Softmax function.

8. The dynamic residual hybrid expert low-light image enhancement method according to claim 6 or 7, characterized in that, The improved denoising network is an asymmetric multi-scale diffusion denoising network, which includes a network layer with a hybrid expert module and a network layer without a hybrid expert module. The network layer with the hybrid expert module is set in the bottleneck layer and the medium-low resolution layer after downsampling, and is used to perform differential feature recovery on dark light areas, noisy areas or structurally missing areas. The network layer without the hybrid expert module is set in the high-resolution layer and the shallow feature layer, and is used to perform basic denoising and detail preservation processing on bright areas and texture fidelity areas. The network layer with the hybrid expert module and the network layer without the hybrid expert module are arranged asymmetrically in a multi-scale hierarchy from shallow to deep.

9. The dynamic residual hybrid expert low-light image enhancement method according to any one of claims 1 to 8, characterized in that, The method further includes: The training process of the dynamic residual hybrid expert low-light image enhancement method includes applying constraint losses to multiple preset key computing nodes. Each constraint loss is calculated based on the difference between the output features and target features of the corresponding node, and the weighted sum of all constraint losses is used as the total loss function to jointly optimize the model parameters.

10. A dynamic residual hybrid expert low-light image enhancement device, characterized in that, include: The memory is configured to store instructions; The processor is configured to retrieve the instructions from the memory and, when executing the instructions, to implement the dynamic residual mixing expert low-light image enhancement method according to any one of claims 1 to 9.

Images