An unsupervised low-light image enhancement method based on deep guided expansion network

Abstract

This invention discloses an unsupervised low-light image enhancement method based on a depth-guided unfolding network, comprising the following steps: constructing a set of low-light domain images and a set of normal-light domain images as learning sample sources; inputting a low-light image, a pre-trained monocular depth estimation model performs forward prediction on the low-light image to obtain a corresponding depth map, and normalizing the depth map to keep its numerical distribution within a uniform range; performing initial decomposition processing on the input low-light image to obtain initial reflectivity and initial illumination for subsequent unfolding optimization; inputting the low-light image, depth map, initial reflectivity, and initial illumination together into a depth unfolding optimization module, multiplying the obtained reflectivity component and illumination component pixel by pixel to output the enhanced image. This invention retains the physical interpretability of Retinex decomposition, combines the powerful modeling capabilities of deep learning, and further introduces scene geometric information for spatial adaptive adjustment.

Description

Technical Field

[0001] This invention relates to the field of computer vision and image processing technology, and in particular to an unsupervised low-light image enhancement method based on a deep guided unfolding network. Background Technology

[0002] Low-light image enhancement aims to recover images with higher visual quality and stronger machine perceptibility from degraded images suffering from insufficient lighting, amplified noise, color distortion, and reduced local contrast. It has significant application value in scenarios such as security monitoring, intelligent terminal imaging, autonomous driving, and robot vision. Existing low-light image enhancement methods can be broadly classified into supervised learning methods and unsupervised learning methods. Supervised learning methods typically rely on strictly paired low-light and normal-light images as training samples. However, in real-world scenarios, obtaining high-quality, strictly registered paired samples is challenging due to factors such as exposure differences, viewpoint shifts, dynamic object motion, variations in noise distribution, and differences in image processing procedures within imaging devices. Therefore, models trained based on paired samples often suffer from insufficient generalization ability and weak adaptability to real-world scenarios.

[0003] On the other hand, traditional Retinex theory posits that an image can be decomposed into reflectance and illumination components. Reflectance describes the inherent color and texture information of the scene, while the illumination component describes variations in external lighting. This theory possesses strong physical interpretability and is therefore widely used in low-light image enhancement tasks. However, traditional Retinex models typically rely on manually designed priors and optimization algorithms, resulting in limited expressive power and difficulty in effectively handling complex non-uniform degradation, noise amplification, and edge preservation problems in real-world scenes.

[0004] Furthermore, while many existing unsupervised enhancement methods have freed themselves from the dependence on strictly paired samples, they often lack sufficient physical constraints and geometric priors during the enhancement process, easily leading to problems such as over-enhancement, color shift, texture distortion, and blurred boundaries. They also struggle to simultaneously achieve brightness restoration, visual naturalness, and structural consistency. To overcome the shortcomings of existing technologies, there is an urgent need for an unsupervised low-light image enhancement method that can retain the physical interpretability of Retinex decomposition, combine the powerful modeling capabilities of deep learning, and further incorporate scene geometric information for spatial adaptive adjustment. This would improve the enhancement effect and generalization ability of the model in real-world, complex scenes. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and to provide an unsupervised low-light image enhancement method that can retain the physical interpretability of Retinex decomposition, combine the powerful modeling capabilities of deep learning, and further introduce scene geometric information for spatial adaptive adjustment, so as to improve the enhancement effect and generalization ability of the model in real complex scenes.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solution:

[0007] An unsupervised low-light image enhancement method based on a deep guided unfolding network includes the following steps:

[0008] (1) Construct a set of low-light domain images and a set of normal-light domain images as sources of learning samples;

[0009] (2) Input a low-light image, pre-trained monocular depth estimation model performs forward prediction on the low-light image to obtain the corresponding depth map, and normalize the depth map so that its numerical distribution remains within a uniform range.

[0010] (3) Perform initialization decomposition processing on the input low-light image to obtain the initial reflectance and initial illumination for subsequent unfolding optimization;

[0011] (4) Input the low-light image, depth map, initial reflectance and initial illumination into the depth expansion optimization module, multiply the obtained reflectance component and illumination component pixel by pixel, and output the enhanced image.

[0012] Furthermore, in step (1), the low-light domain samples are mainly from the LOL dataset and a collection of real low-light images commonly used in related low-light enhancement studies, while the normal-light domain samples are from the RAISE original image dataset and some high dynamic range image sources.

[0013] Furthermore, in step (2), the monocular depth estimation model is Depth Anything V2, and the pre-trained monocular depth estimation model is used to obtain the original depth prediction result corresponding to the input image. Then, it is normalized and calculated as follows:

[0014] (1)

[0015] in, This represents the original depth map output by the monocular depth estimation model. This represents the normalized depth map. The constant is minimal to avoid zero in the denominator. The processed depth map does not participate in the training parameter update but is used as subsequent depth-guided information to construct depth-related smoothing constraints, illumination trend constraints, and conditional guided features. After obtaining the depth map, the low-light image enhancement problem is modeled as a depth-guided Retinex decomposition optimization problem, with the following energy function:

[0016] (2)

[0017] in, This is a data fidelity term used to constrain the recombination of the reflectance and illumination components to approximate the original low-light image. Φ D (R|D) represents the depth-adaptive smoothing term of reflectivity, Ψ D (L|D) represents the depth-adaptive smoothing term of the lighting. (L|D) represents the depth trend constraint term of the lighting, and α, β, and ρ represent the weight coefficients of the above constraint term, respectively. The reflectance and depth adaptive smoothing terms of the lighting are as follows:

[0018] (3)

[0019] (4)

[0020] Where i and j represent the positions of adjacent pixels, R i and R j L represents the reflectance value at adjacent locations. i and L j w represents the illumination value at adjacent locations. ij (D) represents the depth consistency weight calculated from the depth map. The depth consistency weight is calculated as follows:

[0021] (5)

[0022] Among them, D i and D j σ represents the depth values ​​at adjacent pixel positions i and j, respectively. d To control the hyperparameter of the weight decay rate, when D i With D j When approaching, w ij (D) is close to 1, indicating that the region should maintain strong smoothness and consistency. When the difference between the two is large, w ij (D) Rapid decrease indicates that there may be structural boundaries or shading relationships at this location, and the cross-regional smoothing intensity should be reduced.

[0023] Furthermore, in step (3), the initialization decomposition module first estimates a coarse illumination map based on the color channel information of the input image, and obtains the initial illumination value through smoothing. Then, it calculates the coarse initial reflectance value based on the relationship between the low-illuminance image and the initial illumination value. The calculation method is as follows:

[0024] (6)

[0025] in, This indicates that the input image is a low-light image. Represents the reflectivity component, which mainly includes the scene's inherent color, texture, and edge information. The illumination component is represented by ⊙, which mainly describes the spatial light intensity distribution. ⊙ represents pixel-wise multiplication. Finally, the initial reflectance and initial illumination values ​​are corrected by a lightweight convolution thinning branch to obtain the initial reflectance and initial illumination values ​​for subsequent unfolding optimization.

[0026] Furthermore, in step (4), the deep decomposition optimization module includes a reflectance auxiliary variable analysis and update unit, an illumination auxiliary variable analysis and update unit, a reflectance near-end unit, and an illumination near-end unit. The reflectance auxiliary variable analysis and update unit completes the closed-loop update of the reflectance auxiliary variable based on the current low-illuminance observation, reflectance estimation, and illumination auxiliary variable. The calculation method is as follows:

[0027] (7)

[0028] Among them, P k R represents the reflectance auxiliary variable obtained in the k-th expansion stage. k−1 Q represents the reflectivity component output in the previous stage. k−1 Indicates the lighting auxiliary variable corresponding to the previous stage, µ k Let represent the reflectance penalty coefficient for stage k, where 1 indicates a full-1 tensor. Subsequently, the lighting auxiliary variables are analytically updated. The analytical update unit for lighting auxiliary variables completes a closed-form update of the lighting auxiliary variables based on the current low-light observations, lighting estimates, and reflectance auxiliary variables. The calculation method is as follows:

[0029] (8)

[0030] Among them, Q k L represents the lighting auxiliary variable obtained in the k-th expansion stage. k−1 λ represents the illumination component output from the previous stage. k This represents the illumination penalty coefficient for the k-th stage, and c represents the color channel index. This represents the value of the c-th color channel in a low-light image. This represents the value of the reflectance auxiliary variable in the c-th color channel. Since the illumination component is usually represented in single-channel form, this update uses all three RGB channels to constrain the illumination estimation results. (This is followed by a seemingly unrelated sentence about P.) k and Q k After the parsing update, the reflectance component is further updated through depth-guided learnable proximal units. The reflectance proximal units take the reflectance auxiliary variable and illumination auxiliary variable of the current stage as the main input, and introduce the guided features extracted from the depth map. The calculation method is as follows:

[0031] (9)

[0032] Among them, R k G represents the reflectance component updated in the k-th stage. R (∙) represents the near-end mapping of reflectance, θ R ∆R represents the learnable parameters of the reflectivity near-end network. k Net represents the reflectivity residual compensation term. R (∙) represents the convolutional network used to learn the reflectance residuals, and ϕ(D) represents the depth-guided features extracted from the depth map D. Finally, illumination near-end correction is performed. The illumination near-end unit takes the current illumination auxiliary variable and the spatial guidance information constructed from the depth map as input to further correct the illumination components. The calculation method is as follows:

[0033] (10)

[0034] Among them, L k G represents the lighting component updated in the k-th stage. L (∙) represents the near-end illumination mapping, θ L ∆L represents the learnable parameters of the near-end network of the lighting system. k Net represents the lighting residual compensation term. L (∙) represents the convolutional network used to learn the illumination residuals, ψ(D) represents the spatial weights or modulation information constructed from the depth map, f(D) represents the depth-induced illumination trend function, and ∇f(D) represents the gradient direction of the illumination trend. In summary, the four update steps within a single unfolding stage can be summarized as follows:

[0035] (11)

[0036] Among them, F P and F Q To parse the closed-form update operator and preserve data fidelity semantics during Retinex optimization, G R and G LThis is a deep-guided learnable proximal mapping used to supplement traditional optimization models in areas such as texture correction, noise suppression, and illumination recovery, which are difficult to characterize manually. Finally, the reflectivity component R obtained in the last unfolding stage is... T and illumination component L T Pixel-by-pixel multiplication yields the enhanced low-light image, calculated as follows:

[0037] (12)

[0038] in, R represents the final enhanced image. T and L T These represent the reflectivity component and illumination component output in the Tth unfolding stage, respectively.

[0039] It also includes a computer device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, it completes an enhanced low-light image based on an unsupervised low-light image enhancement method based on a deep guided unfolding network.

[0040] It also includes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs an enhanced low-light image based on an unsupervised low-light image enhancement method using a deep guided unfolding network.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] 1. This invention introduces an external depth prior in low-light image enhancement tasks, enabling the pre-trained monocular depth estimation model to achieve spatial adaptive restoration using scene geometric information. This allows for more effective differentiation between flat regions and structural boundary regions in non-uniformly degraded scenes, improving the hierarchy and structural preservation of the enhancement results. The physically meaningful Retinex decomposition optimization process is transformed into a trainable multi-stage network structure, giving the model both the interpretability of traditional optimization methods and the high expressive power of deep learning methods, thus improving model stability and controllability. Through the cooperation of the initialization decomposition module and the depth unfolding optimization module, coarse decomposition results are obtained first, followed by stage-by-stage refinement of reflectivity and illumination components, thereby reducing the difficulty of direct end-to-end restoration and improving the numerical stability during network training and inference.

[0043] 2. This invention constructs an unsupervised adversarial learning framework that enables model training without requiring strict pairing of low-light and normal-light images, reducing data acquisition and annotation costs and enhancing the method's generalization ability in real-world complex scenes. It exhibits good overall performance in brightness restoration, color correction, texture preservation, and structural consistency, providing higher-quality input images for subsequent high-level vision tasks such as object detection, segmentation, and recognition. Attached Figure Description

[0044] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0045] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0046] like Figure 1 As shown, an unsupervised low-light image enhancement method based on a deep guided unfolding network includes the following steps:

[0047] (1) Construct a set of low-light domain images and a set of normal-light domain images as sources of learning samples;

[0048] (2) Input a low-light image, pre-trained monocular depth estimation model performs forward prediction on the low-light image to obtain the corresponding depth map, and normalize the depth map so that its numerical distribution remains within a uniform range.

[0049] (3) Perform initialization decomposition processing on the input low-light image to obtain the initial reflectance and initial illumination for subsequent unfolding optimization;

[0050] (4) Input the low-light image, depth map, initial reflectance and initial illumination into the depth expansion optimization module, multiply the obtained reflectance component and illumination component pixel by pixel, and output the enhanced image.

[0051] The training and validation process in this embodiment employs an unpaired data-driven approach. Specifically, during the training phase, low-light and normal-light image sets are constructed as unsupervised learning sample sources. The low-light samples primarily come from the LOL dataset and commonly used real low-light image sets in related low-light enhancement research, while the normal-light samples come from the RAISE original image dataset and some high dynamic range image sources. To reduce the ambiguity of the training distribution, the training images are first uniformly converted to PNG format and scaled to a fixed resolution of 600×400. Further data augmentation operations such as random cropping and horizontal flipping are employed during training. This embodiment was implemented and trained on an Ubuntu 20.04 system, using an NVIDIA RTX 4090 graphics processor and the PyTorch framework. The AdamW optimizer was selected during training, with a batch size of 4 and an initial learning rate of 1×10⁻⁶. -4 The learning rate is dynamically adjusted using a cosine decay strategy, while a fixed random seed is used to ensure the reproducibility of experimental results.

[0052] First, a low-light image is input. To incorporate scene geometry priors, a pre-trained monocular depth estimation model, Depth Anything V2, is used to perform forward prediction on this low-light image, obtaining the original depth prediction result corresponding to the input image. Then, it is normalized and calculated as follows:

[0053] (1)

[0054] in, This represents the original depth map output by the monocular depth estimation model. This represents the normalized depth map. The value is a very small constant used to avoid the denominator being zero. The processed depth map does not participate in the training parameter update, but is used as subsequent depth guidance information to construct depth-related smoothing constraints, lighting trend constraints, and conditional guidance features.

[0055] After obtaining the depth map, the low-light image enhancement problem is modeled as a depth-guided Retinex decomposition optimization problem, with the following energy function:

[0056] (2)

[0057] in, This is a data fidelity term used to constrain the recombination of the reflectance and illumination components to approximate the original low-light image. Φ D (R|D) represents the depth-adaptive smoothing term of reflectivity, Ψ D (L|D) represents the depth-adaptive smoothing term of the lighting. (L|D) represents the depth trend constraint term of the lighting, and α, β and ρ represent the weight coefficients of the above constraint term, respectively.

[0058] The depth-adaptive smoothing terms for reflectivity and illumination are as follows:

[0059] (3)

[0060] (4)

[0061] Where i and j represent the positions of adjacent pixels, R i and R j L represents the reflectance value at adjacent locations. i and L j w represents the illumination value at adjacent locations. ij (D) represents the depth consistency weight calculated from the depth map. The purpose of the above constraints is to strengthen the smoothing constraint when the depth values ​​of two adjacent pixels are close, making the reflectivity and illumination of the same object or the same spatial plane more continuous; when the depth difference between two adjacent pixels is large, the smoothing constraint is weakened to avoid over-smoothing at object boundaries or occlusion locations.

[0062] The deep consistency weight is calculated as follows:

[0063] (5)

[0064] Among them, D i and D j σ represents the depth values ​​at adjacent pixel positions i and j, respectively. d To control the hyperparameter of the weight decay rate, when D i With D j When approaching, w ij (D) is close to 1, indicating that the region should maintain strong smoothness and consistency. When the difference between the two is large, w ij (D) Rapid decrease indicates that there may be structural boundaries or shading relationships at this location, and the cross-regional smoothing intensity should be reduced.

[0065] The input low-light image undergoes initialization decomposition processing. The initialization decomposition module first estimates a coarse illumination map based on the color channel information of the input image and obtains initial illumination values ​​through smoothing. Then, it calculates a coarse initial reflectance value based on the relationship between the low-light image and the initial illumination value. According to Retinex decomposition theory, the low-light image can be represented as a pixel-wise product of the reflectance component and the illumination component, calculated as follows:

[0066] (6)

[0067] in, This indicates that the input image is a low-light image. Represents the reflectivity component, which mainly includes the scene's inherent color, texture, and edge information. The illumination component represents the spatial light intensity distribution, and ⊙ denotes pixel-wise multiplication. Finally, the initial reflectance and initial illumination values ​​are corrected through lightweight convolutional thinning branches to obtain the initial reflectance and initial illumination values ​​used for subsequent unfolding and optimization. This initialization decomposition process ensures that subsequent network iterations do not start from random states, but rather from states with clear physical meaning and relatively stable numerical distributions.

[0068] The low-light image, depth map, initial reflectance, and initial illumination are jointly input into the depth unrolling optimization module, which includes a reflectance auxiliary variable analysis and update unit, an illumination auxiliary variable analysis and update unit, a reflectance near-end unit, and an illumination near-end unit.

[0069] The reflectance auxiliary variable analysis and update unit completes a closed-loop update of the reflectance auxiliary variable based on the current low-light observations, reflectance estimates, and illumination auxiliary variables. The calculation method is as follows:

[0070] (7)

[0071] Among them, P k R represents the reflectance auxiliary variable obtained in the k-th expansion stage. k−1 Q represents the reflectivity component output in the previous stage. k−1 Indicates the lighting auxiliary variable corresponding to the previous stage, µ k Let represent the reflectance penalty coefficient for the k-th stage, and 1 denote the all-1 tensor. This formula indicates that the update of the reflectance auxiliary variable is simultaneously constrained by the consistency of the input image data and the reflectance estimation of the previous stage, thereby ensuring that the update result has good physical consistency and stage continuity.

[0072] Subsequently, the lighting auxiliary variables are updated analytically. The lighting auxiliary variable analytical update unit completes the closed-loop update of the lighting auxiliary variables based on the current low-illuminance observation, illumination estimate, and reflectance auxiliary variable. The calculation method is as follows:

[0073] (8)

[0074] Among them, Q k L represents the lighting auxiliary variable obtained in the k-th expansion stage. k−1 λ represents the illumination component output from the previous stage. k This represents the illumination penalty coefficient for the k-th stage, and c represents the color channel index. This represents the value of the c-th color channel in a low-light image. This represents the value of the reflectance auxiliary variable in the c-th color channel. Since illumination components are usually represented in single-channel form, this update uses all three RGB channels to constrain the illumination estimation results.

[0075] After completing P k and Q k After the parsing update, the reflectance component is further updated through depth-guided learnable proximal units. The reflectance proximal units take the reflectance auxiliary variable and illumination auxiliary variable of the current stage as the main input, and introduce the guided features extracted from the depth map. The calculation method is as follows:

[0076] (9)

[0077] Among them, R k G represents the reflectance component updated in the k-th stage. R (∙) represents the near-end mapping of reflectance, θ R ∆R represents the learnable parameters of the reflectivity near-end network. k Net represents the reflectivity residual compensation term. R (∙) represents the convolutional network used to learn the reflectance residual, and ϕ(D) represents the depth-guided feature extracted from the depth map D. This formula indicates that the reflectance proximal unit does not directly regenerate reflectance, but rather updates the reflectance in the parsed result P. k Based on this, residual correction values ​​are learned to enhance texture recovery capabilities and suppress noise interference.

[0078] Finally, illumination near-end correction is performed. The illumination near-end unit takes the current illumination auxiliary variable and spatial guidance information constructed from the depth map as input to further correct the illumination components. The calculation method is as follows:

[0079] (10)

[0080] Among them, L k G represents the lighting component updated in the k-th stage. L (∙) represents the near-end illumination mapping, θ L ∆L represents the learnable parameters of the near-end network of the lighting system. k Net represents the lighting residual compensation term. L (∙) denotes the convolutional network used to learn the lighting residuals, ψ(D) represents the spatial weights or modulation information constructed from the depth map, f(D) represents the depth-induced lighting trend function, and ∇f(D) represents the gradient direction of the lighting trend. This equation indicates that the lighting update depends not only on the current lighting auxiliary variable Q. k It also utilizes depth information to guide lighting to remain smooth within the same spatial area and to reduce excessive diffusion at abrupt changes in depth.

[0081] In summary, the four update steps within a single unfolding phase can be summarized as follows:

[0082] (11)

[0083] Among them, F P and F Q To parse the closed-form update operator and preserve data fidelity semantics during Retinex optimization, G R and G L This is a deep-guided learnable proximal mapping used to supplement traditional optimization models in areas such as texture correction, noise suppression, and illumination recovery, which are difficult to characterize manually. Multiple unfolding stages are cascaded sequentially to achieve stepwise optimization of reflectivity and illumination components from coarse to fine.

[0084] Finally, the reflectivity component R obtained in the last unfolding stage T and illumination component L T Pixel-by-pixel multiplication yields the enhanced low-light image, calculated as follows:

[0085] (12)

[0086] in, R represents the final enhanced image. T and L T These represent the reflectivity component and illumination component output in the Tth unfolding stage, respectively.

[0087] In this embodiment, the network structure remains consistent across multiple unfolding stages, and the main parameters of the reflectance and illumination near-end update networks can be shared between stages. The penalty control parameters in each stage can be adjusted according to the iteration sequence. This design maintains model compactness while simulating the gradually tightening constraints characteristic of traditional optimization processes, allowing early stages to prioritize overall brightness and structure restoration, and later stages to prioritize detail enhancement and artifact suppression. After all unfolding stages are completed, the final reflectance component is multiplied pixel-by-pixel with the illumination component to output the enhanced image. Because this output is simultaneously influenced by data consistency, depth geometric consistency, and learned prior constraints, it effectively maintains color fidelity and structural clarity while improving overall brightness.

[0088] To achieve training without paired data, this embodiment further constructs an unsupervised adversarial learning framework. Specifically, the aforementioned deep unfolded network is used as the generator, and a lightweight convolutional discriminator is introduced. During training, low-light images are sampled from the low-light image dataset as input to the generator, and real normal-light images are sampled from the normal-light image dataset as real samples for the discriminator. The generator is responsible for mapping low-light images to the normal-light style domain, and the discriminator is responsible for distinguishing between enhanced images and real normal-light images. Through alternating optimization of the generator and discriminator, the enhancement result gradually approximates the visual distribution of real normal-light images.

[0089] In addition to adversarial learning, this embodiment also sets up multiple auxiliary losses to jointly constrain the network training process. Among them, the Retinex consistency constraint is used to ensure the consistency between the enhancement result and the reflectivity-illumination decomposition relationship; the perceptual consistency constraint is used to improve the naturalness of the enhancement result in the high-level semantic feature space; and the depth-guided smoothing constraint is used to ensure that the illumination restoration process can maintain coordination with changes in scene geometry. Through the joint optimization of the above multiple constraints, the network can balance physical rationality, visual naturalness, and structural consistency under unsupervised conditions.

Claims

1. An unsupervised low-light image enhancement method based on a deep guided unfolding network, characterized in that: Includes the following steps: (1) Construct a set of low-light domain images and a set of normal-light domain images as sources of learning samples; (2) Input a low-light image, pre-trained monocular depth estimation model performs forward prediction on the low-light image to obtain the corresponding depth map, and normalize the depth map so that its numerical distribution remains within a uniform range. (3) Perform initialization decomposition processing on the input low-light image to obtain the initial reflectance and initial illumination for subsequent unfolding optimization; (4) Input the low-light image, depth map, initial reflectance and initial illumination into the depth expansion optimization module, multiply the obtained reflectance component and illumination component pixel by pixel, and output the enhanced image.

2. The unsupervised low-light image enhancement method based on a deep guided unfolding network according to claim 1, characterized in that: In step (1), the low-light domain samples mainly come from the LOL dataset and a collection of real low-light images commonly used in related low-light enhancement studies, while the normal-light domain samples come from the RAISE original image dataset and some high dynamic range image sources.

3. The unsupervised low-light image enhancement method based on a deep guided unfolding network according to claim 1, characterized in that: In step (2), the monocular depth estimation model is Depth Anything V2. The pre-trained monocular depth estimation model is used to obtain the original depth prediction result corresponding to the input image. Then, it is normalized and calculated as follows: （1） in, This represents the original depth map output by the monocular depth estimation model. This represents the normalized depth map. The constant is minimal to avoid zero in the denominator. The processed depth map does not participate in the training parameter update but is used as subsequent depth-guided information to construct depth-related smoothing constraints, illumination trend constraints, and conditional guided features. After obtaining the depth map, the low-light image enhancement problem is modeled as a depth-guided Retinex decomposition optimization problem, with the following energy function: （2） in, This is a data fidelity term used to constrain the recombination of the reflectance and illumination components to approximate the original low-light image. Φ D (R|D) represents the depth-adaptive smoothing term of reflectivity, Ψ D (L|D) represents the depth-adaptive smoothing term of the lighting. (L|D) represents the depth trend constraint term of the lighting, and α, β, and ρ represent the weight coefficients of the above constraint term, respectively. The reflectance and depth adaptive smoothing terms of the lighting are as follows: （3） （4） Where i and j represent the positions of adjacent pixels, R i and R j L represents the reflectance value at adjacent locations. i and L j w represents the illumination value at adjacent locations. ij (D) represents the depth consistency weight calculated from the depth map. The depth consistency weight is calculated as follows: （5） Among them, D i and D j σ represents the depth values ​​at adjacent pixel positions i and j, respectively. d To control the hyperparameter of the weight decay rate, when D i With D j When approaching, w ij (D) is close to 1, indicating that the region should maintain strong smoothness and consistency. When the difference between the two is large, w ij (D) Rapid decrease indicates that there may be structural boundaries or shading relationships at this location, and the cross-regional smoothing intensity should be reduced.

4. The unsupervised low-light image enhancement method based on a deep guided unfolding network according to claim 1, characterized in that: In step (3), the initialization decomposition module first estimates a coarse illumination map based on the color channel information of the input image, and obtains the initial illumination value through smoothing. Then, it calculates the coarse initial reflectance value based on the relationship between the low-illuminance image and the initial illumination value. The calculation method is as follows: （6） in, This indicates that the input image is a low-light image. Represents the reflectivity component, which mainly includes the scene's inherent color, texture, and edge information. The illumination component is represented by ⊙, which mainly describes the spatial light intensity distribution. ⊙ represents pixel-wise multiplication. Finally, the initial reflectance and initial illumination values ​​are corrected by a lightweight convolution thinning branch to obtain the initial reflectance and initial illumination values ​​for subsequent unfolding optimization.

5. The unsupervised low-light image enhancement method based on a deep guided unfolding network according to claim 1, characterized in that: In step (4), the deep decomposition optimization module includes a reflectance auxiliary variable analysis and update unit, an illumination auxiliary variable analysis and update unit, a reflectance near-end unit, and an illumination near-end unit. The reflectance auxiliary variable analysis and update unit completes the closed-loop update of the reflectance auxiliary variable based on the current low-illuminance observation, reflectance estimation, and illumination auxiliary variable. The calculation method is as follows: （7） Among them, P k R represents the reflectance auxiliary variable obtained in the k-th expansion stage. k−1 Q represents the reflectivity component output in the previous stage. k−1 Indicates the lighting auxiliary variable corresponding to the previous stage, µ k Let represent the reflectance penalty coefficient for stage k, where 1 indicates a full-1 tensor. Subsequently, the lighting auxiliary variables are analytically updated. The analytical update unit for lighting auxiliary variables completes a closed-form update of the lighting auxiliary variables based on the current low-light observations, lighting estimates, and reflectance auxiliary variables. The calculation method is as follows: （8） Among them, Q k L represents the lighting auxiliary variable obtained in the k-th expansion stage. k−1 λ represents the illumination component output from the previous stage. k This represents the illumination penalty coefficient for the k-th stage, and c represents the color channel index. This represents the value of the c-th color channel in a low-light image. This represents the value of the reflectance auxiliary variable in the c-th color channel. Since the illumination component is usually represented in single-channel form, this update uses all three RGB channels to constrain the illumination estimation results. (This is followed by a seemingly unrelated sentence about P.) k and Q k After the parsing update, the reflectance component is further updated through depth-guided learnable proximal units. The reflectance proximal units take the reflectance auxiliary variable and illumination auxiliary variable of the current stage as the main input, and introduce the guided features extracted from the depth map. The calculation method is as follows: （9） Among them, R k G represents the reflectance component updated in the k-th stage. R (∙) represents the near-end mapping of reflectance, θ R ∆R represents the learnable parameters of the reflectivity near-end network. k Net represents the reflectivity residual compensation term. R (∙) represents the convolutional network used to learn the reflectance residuals, and ϕ(D) represents the depth-guided features extracted from the depth map D. Finally, illumination near-end correction is performed. The illumination near-end unit takes the current illumination auxiliary variable and the spatial guidance information constructed from the depth map as input to further correct the illumination components. The calculation method is as follows: （10） Among them, L k G represents the lighting component updated in the k-th stage. L (∙) represents the near-end illumination mapping, θ L ∆L represents the learnable parameters of the near-end network of the lighting system. k Net represents the lighting residual compensation term. L (∙) represents the convolutional network used to learn the illumination residuals, ψ(D) represents the spatial weights or modulation information constructed from the depth map, f(D) represents the depth-induced illumination trend function, and ∇f(D) represents the gradient direction of the illumination trend. In summary, the four update steps within a single unfolding stage can be summarized as follows: （11） Among them, F P and F Q To parse the closed-form update operator and preserve data fidelity semantics during Retinex optimization, G R and G L This is a deep-guided learnable proximal mapping used to supplement traditional optimization models in areas such as texture correction, noise suppression, and illumination recovery, which are difficult to characterize manually. Finally, the reflectivity component R obtained in the last unfolding stage is... T and illumination component L T Pixel-by-pixel multiplication yields the enhanced low-light image, calculated as follows: （12） in, R represents the final enhanced image. T and L T These represent the reflectivity component and illumination component output in the Tth unfolding stage, respectively.

6. A computer device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, the following features are provided: An unsupervised low-light image enhancement method based on a deep guided unfolding network, as described in any one of claims 1-5, is used to complete the enhanced low-light image.

7. A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the following features are provided: An unsupervised low-light image enhancement method based on a deep guided unfolding network, as described in any one of claims 1-5, is used to complete the enhanced low-light image.

Images