A low-light image enhancement method and device based on physical illumination model constraint, electronic equipment and storage medium

Abstract

The application discloses a low-light image enhancement method and device based on physical illumination model constraint, electronic equipment and storage medium, and the method comprises the steps of acquiring a low-light image, decomposing the low-light image into a specular reflection component, a diffuse reflection component and an indirect illumination component according to a direct diffuse reflection-specular reflection-indirect illumination model; initializing the specular reflection component by using a surface normal map of the low-light image and a pre-trained specular inverse rendering module; setting a full-zero pixel matrix consistent with the low-light image to be enhanced as an initial indirect illumination component; constructing a physically guided depth unfolding network with the energy minimization of the model as the target, and iteratively updating the diffuse reflection component, the indirect illumination component and the specular reflection component alternately; after the iteration is terminated, the finally obtained diffuse reflection component, indirect illumination component and specular reflection component are superimposed pixel by pixel to generate and output an enhanced low-light image. Through the above setting, the physical authenticity, the detail integrity and the visual naturalness of the low-light image are enhanced.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to a low-light image enhancement method, apparatus, electronic device and storage medium based on physical illumination model constraints. Background Technology

[0002] Images captured under low-light conditions commonly suffer from insufficient brightness, increased noise, low contrast, and loss of detail, severely impacting subjective visual experience and the performance of subsequent high-level visual tasks such as object detection and semantic segmentation. Traditional low-light enhancement methods, such as histogram equalization and gamma correction, only increase brightness by adjusting pixel grayscale values, without considering local illumination differences in the image, easily leading to loss of detail and amplified noise. Early methods based on Retinex theory relied on manual prior knowledge, exhibiting poor adaptability in complex scenes, and the iterative optimization process was computationally expensive. In recent years, deep learning methods have significantly improved enhancement effects, but purely data-driven models are mostly "black box" structures, lacking physical interpretability. When dealing with complex optical phenomena such as specular highlights, they are prone to problems such as color deviation, overly sharp textures, or overexposure in highlight areas.

[0003] While existing methods for low-light enhancement based on depth unfolding achieve rapid end-to-end enhancement, most remain confined to the Retinex multiplicative decomposition framework, lacking sufficient illumination modeling and failing to introduce strict geometric and optical constraints. Methods based on a two-color reflection model, while capable of decomposing diffuse and specular reflection components, treat specular highlights as data-driven residual features, failing to truly incorporate physical constraints, resulting in a lack of physical consistency in highlight region recovery. Furthermore, existing methods generally neglect crucial multiple scattering and indirect illumination effects in low-light environments, failing to fully characterize the illumination distribution of real-world scenes. This leads to enhancement results prone to defects such as loss of detail in dark areas, uneven illumination, and specular distortion. Summary of the Invention

[0004] To address the shortcomings mentioned above, this application provides a low-light image enhancement method, apparatus, electronic device, and storage medium based on physical illumination model constraints.

[0005] Firstly, to achieve the above objectives, this application provides a low-light image enhancement method based on physical illumination model constraints, comprising: The low-light image to be enhanced is obtained, and the low-light image is decomposed into diffuse reflection component, specular reflection component and indirect illumination component according to the direct diffuse reflection-specular reflection-indirect illumination model. The surface normal map corresponding to the low-light image is obtained using a normal estimation network; the low-light image and the surface normal map are subjected to specular inverse rendering to obtain the initial specular reflection component; the all-zero pixel matrix that is consistent with the low-light image to be enhanced is set as the initial indirect lighting component. With the goal of minimizing the energy of the direct diffuse-spectral-indirect lighting model, a physically guided depth unfolding network is constructed. In each iteration stage, the diffuse component, indirect lighting component, and specular component are updated alternately: the low-light image is subtracted from the first iteration. The residual image obtained from the specular reflection component and the indirect illumination component of the first stage is input into the diffuse reflection update network to update the second stage. The diffuse component of the stage; subtracting the first stage from the low-light image. The diffuse component updated in the first stage and the first The residual image obtained from the specular reflection component of the first stage is input into the indirect illumination update network to update the second stage. The indirect illumination component of the stage; combined with the first The diffuse component, indirect illumination component, and the first phase of the update are updated. The specular reflection components of each stage are combined to form an intermediate reconstructed image. This image, combined with the surface normal map, is then processed by the specular inverse rendering module to update the next stage. The mirror reflection component of the stage; repeat the iterative update process until the energy function converges or reaches the preset termination condition; After the iteration terminates, the final diffuse reflection component, specular reflection component, and indirect illumination component are superimposed pixel by pixel to generate and output the enhanced low-light image.

[0006] Preferably, the energy minimization objective function formula for the direct diffuse reflection-specular reflection-indirect illumination model is as follows: ; In the formula: This is the diffuse reflection component; This is the specular reflection component; This is the indirect illumination component; Low-light image; This is a diffuse regularization term used to constrain the piecewise smoothness and texture properties of the diffuse component; ; These are the weight parameters.

[0007] Preferably, the iterative update of the diffuse reflection component, indirect illumination component, and specular reflection component, with the goal of minimizing the energy of the direct diffuse reflection-specular reflection-indirect illumination model, includes: The formula for iteratively updating the diffuse reflection component is as follows: ; The formula for iteratively updating the indirect illumination component is as follows: ; The formula for iteratively updating the specular reflection component is as follows: ; In the formula: This represents the current iteration number; For the first The diffuse component after the next update; The next updated specular reflection components; The next update of the indirect lighting component.

[0008] Preferably, the loss function formula for the direct diffuse reflection-specular reflection-indirect illumination model is as follows: ; In the formula: To enhance the L1 loss function of the image; The structural similarity loss function; For the sparse loss function of the mirror component; The constraint loss function for the mirror inverse rendering module; These are the weighting coefficients.

[0009] Preferably, the constraint loss function of the mirror inverse rendering module includes an illumination consistency loss function and an illumination robustness loss function, as shown in the following formula: ; ; ; In the formula: These are the weighting coefficients; The specular reflection intensity coefficient is the predicted value under high light intensity conditions; The specular reflection intensity coefficient is the predicted value under low light intensity conditions; The specular index is the result of prediction under high light intensity conditions. The specular index is the result of prediction under low light intensity conditions; This is a network for estimating mirror parameters, used to predict mirror reflection parameters based on the input image and the direction of illumination; This is the low-light image at the t-th iteration; This is the predicted illumination direction vector; This is a random, small perturbation vector used to simulate the uncertainty of the illumination direction; for Learnable parameters of the network.

[0010] Preferably, the formula for the mirror inverse rendering module is as follows: ; ; In the formula: The mirror intensity coefficient; This is a surface normal diagram; The direction vector of the light; It is a half-range vector; The direction vector is fixed at the viewpoint. This refers to the highlight index.

[0011] Preferably, the specular intensity coefficient and specular index are estimated using a coarse-to-fine strategy, as shown in the following formula: ; ; ; In the formula: The initial mirror intensity coefficient; The initial highlight index; The final optimized mirror intensity coefficient; The final optimized highlight index; Reconstruct the image from the intermediate points; for Learnable parameters of the network; for Learnable parameters of the network.

[0012] Secondly, this application also provides a low-light image enhancement device based on physical illumination model constraints, comprising: The acquisition and initialization module is used to acquire the low-light image to be enhanced, decompose the low-light image into diffuse reflection component, specular reflection component, and indirect lighting component according to the direct diffuse reflection-specular reflection-indirect lighting model; obtain the corresponding surface normal map based on the low-light image using a normal estimation network; perform specular inverse rendering processing on the low-light image and the surface normal map to obtain the initial specular reflection component; and set the all-zero pixel matrix consistent with the low-light image as the initial indirect lighting component. The processing module is used to alternately update the diffuse reflection component, indirect illumination component, and specular reflection component in each iteration stage, with the goal of minimizing the energy of the direct diffuse reflection-spectral reflection-indirect illumination model. This involves subtracting the first iteration from the low-light image. The residual image obtained from the specular reflection component and the indirect illumination component of the first stage is input into the diffuse reflection update network to update the second stage. The diffuse component of the stage; subtract the first from the low-light image. The diffuse component updated in the first stage and the first The residual image obtained from the specular reflection component of the current stage is input into the indirect illumination update network to update the indirect illumination component of the current stage; combined with the first stage... The diffuse reflection, indirect illumination components, and the first phase of the update are as follows: The specular reflection components of each stage are combined to form an intermediate reconstructed image. This image, combined with the surface normal map, is then processed by the specular inverse rendering module to update the next stage. The mirror reflection component of the stage; repeat the iterative update process until the energy function converges or reaches the preset termination condition; The execution module is used to superimpose the final diffuse reflection component, specular reflection component, and indirect illumination component pixel by pixel after the iteration terminates, generating and outputting the enhanced low-light image.

[0013] Thirdly, this application also provides an electronic device, including at least one processing unit and at least one storage unit, wherein the storage unit stores a computer program, and when the program is executed by the processing unit, the processing unit performs the above-described method.

[0014] Fourthly, this application also provides a storage medium storing a computer program executable by an electronic device, which, when run on the electronic device, causes the electronic device to perform the above-described method.

[0015] Compared with the prior art, the beneficial effects of this application are as follows: This application decomposes low-light images into three components using a direct diffuse-specular-indirect lighting model. It then performs specular inverse rendering based on surface normal maps to obtain initial components. With the goal of minimizing energy, it iteratively updates the diffuse, indirect lighting, and specular components. This approach can accurately restore specular highlights, compensate for indirect lighting, and preserve material details while improving brightness. It significantly enhances the physical realism, detail integrity, and visual naturalness of low-light images, meeting the practical needs for high-quality image enhancement in complex low-light scenes. Attached Figure Description

[0016] Figure 1 This is a flowchart of the low-light image enhancement method of this application; Figure 2 This is a comparison of the visual quality of image enhancement on the LOLv1 dataset in this application; Figure 3 This is a comparison of the visual quality of image enhancement on the LOLv2-Real dataset in this application; Figure 4 This is a comparison of the visual quality of image enhancement on the LOLv2-Synthetic dataset in this application. Detailed Implementation

[0017] 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. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] like Figure 1 As shown, this application provides a low-light image enhancement method based on physically constrained lighting models. This method enhances the input low-light image by explicitly modeling diffuse reflection, specular reflection, and indirect lighting components, and constructing a Physically Guided Unfolding Network (PGU-Net). The PGU-Net includes a Physically Guided Unfolding module for decoupling and restoring physical components, a Physically Based Specular Inverse Rendering (PBSIR) module for physically reconstructing specular highlight regions, a Diffuse Reflection Update Network (D-Net) for reconstructing local diffuse reflection features, and an Indirect Lighting Update Network (S-Net) for capturing low-frequency ambient light. The method includes the following steps: The low-light image to be enhanced is obtained, and it is decomposed into specular reflection component, diffuse reflection component and indirect illumination component according to the direct diffuse reflection-specular reflection-indirect illumination model.

[0019] Specifically, the energy minimization objective function formula for the direct diffuse reflection-specular reflection-indirect illumination model is as follows: ; In the formula: This is the diffuse reflection component; This is the specular reflection component; This is the indirect illumination component; Low-light image; This is a diffuse regularization term used to constrain the piecewise smoothness and texture properties of the diffuse component; ; These are the weight parameters.

[0020] By using the above settings, the limitations of the traditional Retinex multiplicative decomposition framework are broken through, and the image is decomposed into three physically interpretable components according to the real optical mechanism, which fully characterizes the direct illumination, specular reflection and multiple scattering effects in light transmission. The surface normal map corresponding to the low-light image is obtained using a normal estimation network; the low-light image and the surface normal map are subjected to specular inverse rendering to obtain the initial specular reflection component; the all-zero pixel matrix that is consistent with the low-light image to be enhanced is set as the initial indirect lighting component. Specifically, a pre-trained normal estimation network is used to extract geometric features from the low-light image, generating a surface normal map N as a geometric prior. A specular inverse rendering module is then used to process the low-light image and the surface normal map to obtain the initial specular reflection components. An all-zero pixel matrix with the same size as the low-light image is used as the initial indirect illumination component. .

[0021] With the above settings, the surface normal is used as a geometric prior to guide the initial estimation of the mirror component, and a stable initialization strategy is adopted to ensure that the subsequent iteration process starts stably, avoiding iteration divergence or enhanced distortion due to unreasonable initial values.

[0022] With the goal of minimizing the energy of the direct diffuse reflection-spectral reflection-indirect illumination model, the diffuse reflection component, specular reflection component, and indirect illumination component are iteratively updated. The iterative update is implemented based on a physically guided deep unfolding network, and an alternating minimization algorithm is used for stage-by-stage optimization. The specific steps are as follows: Subtract the first from the low-light image The residual image obtained from the specular reflection component and the indirect illumination component of the first stage is input into the diffuse reflection update network to update the second stage. The diffuse reflection component of the stage; The update of the diffuse reflection component in stage t satisfies the following formula: ; This formula is mathematically equivalent to performing a near-end mapping on the residual image: ; In the physics-guided deep unfolding network, a diffuse reflection update network (D-Net) based on the U-Net structure is used as a learnable proximal operator. Large convolutional kernel blocks are embedded in the feature bottleneck layer of the D-Net, containing large-size 7×7 deep separable convolutions and cascaded channel attention mechanisms. The D-Net simultaneously extracts material texture and enhances illumination, directly mapping low-light residuals to diffuse reflection components under normal illumination. Its update rule is expressed as: ; Let be the diffuse component after the t-th update; The next updated specular reflection components; The indirect lighting component after the next update; D-Net; D-Net network parameters. By setting these parameters, noise is suppressed while accurately reproducing the inherent texture and details of the object, resulting in stronger structural fidelity for the diffuse component.

[0023] Subtract the first from the low-light image The diffuse component updated in the first stage and the first The residual image obtained from the specular reflection component of the first stage is input into the indirect illumination update network to update the second stage. Indirect light component of the stage; The update of the indirect illumination component in stage t satisfies the following formula: ; This formula is mathematically equivalent to performing a near-end mapping on the residual image: ; An indirect illumination update network (S-Net) is used for updates. A global context block is embedded in the feature bottleneck layer of the S-Net, and a self-attention mechanism is used to calculate the global context dependency to extract global low-frequency indirect illumination changes. The expression is as follows: ; In the formula: S-Net; S-Net network parameters. These settings accurately compensate for indirect lighting in low-light, dark areas, significantly improving image spatial illumination uniformity and the visibility of details in dark areas.

[0024] Combining the updated diffuse reflection, indirect illumination component and the first The specular reflection components of each stage are combined to form an intermediate reconstructed image. Combined with the surface normal map, this image is then updated by the specular inverse rendering module. The specular reflection component of the stage.

[0025] The update of the specular reflection component in stage t satisfies the following formula: ; Based on the mirror inverse rendering module, combined with the differentiable Blinn-Phong physical model and a coarse-to-fine parameter estimation strategy, the final mirror reflection component is generated: ; ; In the formula: The mirror intensity coefficient; This is a surface normal diagram; The direction vector of the light; It is a half-range vector; The direction vector is fixed at the viewpoint. This refers to the highlight index.

[0026] The formulas for specular intensity and specular index are shown below: ; ; ; In the formula: The initial mirror intensity coefficient; The initial highlight index; The final optimized mirror intensity coefficient; The final optimized highlight index; Reconstruct the image from the intermediate points; for Learnable parameters of the network; for Learnable parameters of the network. Through the above settings, specular highlights are explicitly reconstructed based on geometric optics laws and differentiable physical models, solving the problems of overexposure, blurring, and color shift in highlights caused by traditional methods, making the highlight areas natural, sharp, and physically believable.

[0027] Repeat the iterative update process until the energy function converges or the preset termination condition is met; Specifically, the diffuse reflection component, indirect illumination component, and specular reflection component updated in each stage are used as the input components for the next stage, and iterative updates are performed cyclically. The preset termination condition is to reach a fixed number of iterations, such as T=4, the energy function value tends to stabilize, or the difference between adjacent iteration components is less than a preset threshold.

[0028] After the iteration terminates, the final diffuse reflection component, specular reflection component, and indirect illumination component are superimposed pixel by pixel to generate and output the enhanced low-light image. By setting the above parameters, the three physically reasonable components are directly summed and output, ensuring that the enhanced image has uniform global illumination, complete details, low noise, and natural highlights, significantly improving subjective visual quality and objective evaluation indicators.

[0029] Specifically, the final diffuse reflection component after iterative optimization Specular reflection and indirect light component Pixel-by-pixel stacking yields the enhanced image: ; In the formula, This is the enhanced low-light image.

[0030] In the implementation of this application, the loss function formula for the direct diffuse reflection-specular reflection-indirect illumination model is as follows: ; In the formula: To enhance the L1 loss function of the image; The structural similarity loss function; For the sparse loss function of the mirror component; The constraint loss function for the mirror inverse rendering module; represents the weight coefficients. Through the above settings, the training of the physically guided depth unfolding network is constrained from multiple dimensions, including pixel fidelity, structural similarity, physical sparsity, parameter consistency, and illumination robustness. This ensures that the physically guided depth unfolding network still outputs stable and physically consistent enhancement results even in extreme low-light and high-noise environments.

[0031] Furthermore, in order to ensure the final enhanced image Compared with real normal lighting images Pixel-level precision and structural consistency between them, using Norm loss and Structural Similarity (SSIM) loss: ; ; It should be noted that explicit regularization terms for diffuse and indirect illumination components are omitted from the loss function because their corresponding physical priors have been implicitly learned and enforced by data-driven proximal operators (D-Net and S-Net) during the depth unrolling stage. However, for specular highlights, the norm is explicitly preserved to constrain their inherent spatial sparsity. .

[0032] In the implementation of this application, the constraint loss function of the mirror inverse rendering module includes a lighting consistency loss function and a lighting robustness loss function, as shown in the following formula: ; ; ; In the formula: These are the weighting coefficients; The specular reflection intensity coefficient is the predicted value under high light intensity conditions; The specular reflection intensity coefficient is the predicted value under low light intensity conditions; The specular index is the result of prediction under high light intensity conditions. The specular index is the result of prediction under low light intensity conditions; This is a network for estimating mirror parameters, used to predict mirror reflection parameters based on the input image and the direction of illumination; This is the low-light image at the t-th iteration; This is the predicted illumination direction vector; This is a random, small perturbation vector used to simulate the uncertainty of the illumination direction; for Learnable parameters of the network.

[0033] Example 1 The LOLv1 dataset, LOLv2-Real dataset, and LOLv2-Synthetic dataset were used for training and testing; the mirror inverse rendering module was pre-trained on the DiLiGenT photometric stereo dataset.

[0034] Hardware platform: NVIDIA GeForce RTX 3090 GPU; framework: PyTorch; optimizer: Adam; total training epochs: 200; initial learning rate: decay to Number of iterations Loss function weight settings: and They are 0.1 and 0.2 respectively. Set to 0.3, robustness loss weight Set it to 0.2.

[0035] This application compares the image quality with JIEP, RetinexNet, KinD, KinD++, UREtinex-Net, R2RNet, CUE, RSFNet, Retinexformer, QuadPrior, and CIDNet. The evaluation is based on Peak Signal-to-Noise Ratio (PSNR), Structural Similarity (SSIM), Mean Absolute Error (MAE), Feature Similarity (FSIM), and No-Reference Sensing metric NIQE. Higher PSNR, SSIM, and FSIM, or lower MAE and NIQE, indicate better image quality. The experimental results are shown in Table 1.

[0036] Table 1 like Figure 2-4 As shown in the experiment, the proposed method outperforms existing mainstream algorithms in terms of PSNR, FSIM, MAE, etc., and enhances the image with appropriate brightness, rich details, natural highlights, and no obvious noise.

[0037] This application also provides a low-light image enhancement system, including: The acquisition and initialization module is used to acquire the low-light image to be enhanced. Based on the direct diffuse-specular-indirect lighting model, the low-light image is decomposed into diffuse reflection components, specular reflection components, and indirect lighting components. The corresponding surface normal map is obtained from the low-light image using a normal estimation network. The low-light image and the surface normal map are subjected to specular inverse rendering to obtain the initial specular reflection component. The all-zero pixel matrix that is consistent with the low-light image to be enhanced is set as the initial indirect lighting component. The processing module aims to minimize the energy of the direct diffuse-spectral-indirect lighting model, and alternately updates the diffuse component, indirect lighting component, and specular component in each iteration stage: subtracting the low-light image from the... The residual image obtained from the specular reflection component and the indirect illumination component of the current stage is input into the diffuse reflection update network to update the diffuse reflection component of the current stage; the low-light image is subtracted from the first stage. The diffuse component after the phase update and the first The residual image obtained from the specular reflection component of the current stage is input into the indirect illumination update network to update the indirect illumination component of the current stage; combined with the first stage... The diffuse reflection, indirect illumination components, and the first phase of the update are as follows: The specular reflection components of each stage are combined to form an intermediate reconstructed image. This image, combined with the surface normal map, is then processed by the specular inverse rendering module to update the next stage. The mirror reflection component of the stage; repeat the iterative update process until the energy function converges or reaches the preset termination condition; The execution module is used to superimpose the final diffuse reflection component, specular reflection component, and indirect illumination component pixel by pixel after the iteration terminates, generating and outputting the enhanced low-light image.

[0038] This application also provides an electronic device, including at least one processing unit and at least one storage unit, wherein the storage unit stores a computer program, and when the program is executed by the processing unit, the processing unit performs the above-described method.

[0039] This application also provides a storage medium storing a computer program executable by an electronic device, which, when run on the electronic device, causes the electronic device to perform the above-described method.

[0040] The above are merely preferred embodiments of this application and are not intended to limit 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 protection scope of this application.

Claims

1. A low-light image enhancement method based on physical illumination model constraints, characterized in that, include: The low-light image is determined and decomposed into specular reflection component, diffuse reflection component and indirect illumination component according to the direct diffuse reflection-specular reflection-indirect illumination model; Based on the low-light image, the corresponding surface normal map is obtained using a normal estimation network; the low-light image and the surface normal map are subjected to preliminary specular inverse rendering to obtain the initial specular reflection component; a zero matrix with the same size as the low-light image is set as the initial indirect lighting component. With the goal of minimizing the energy of the direct diffuse-specular-indirect illumination model, a physically guided depth unfolding network is constructed. In each iteration stage, the diffuse component, indirect illumination component, and specular component are updated alternately. The low-light image is then subtracted from the first iteration. The residual image obtained from the specular reflection component and the indirect illumination component of the first stage is input into the diffuse reflection update network to update the second stage. The diffuse component of the stage; subtract the first from the low-light image. The diffuse component updated in the first stage and the first The residual image obtained from the specular reflection component of the first stage is input into the indirect illumination update network to update the second stage. Indirect light component of the stage; Combined with the The diffuse component, indirect illumination component, and the first phase of the update are updated. The specular reflection components of each stage are combined to form an intermediate reconstructed image. Combined with the surface normal map, this image is then updated by the specular inverse rendering module. The mirror reflection component of the stage; repeat the iterative update process until the energy function converges or reaches the preset termination condition; After the iteration terminates, the final diffuse reflection component, specular reflection component, and indirect illumination component are superimposed pixel by pixel to generate and output the enhanced low-light image.

2. The low-light image enhancement method according to claim 1, characterized in that, The energy minimization objective function formula for the direct diffuse reflection-specular reflection-indirect illumination model is as follows: ； In the formula: This is the diffuse reflection component; This is the specular reflection component; This is the indirect illumination component; Low-light image; This is a diffuse regularization term used to constrain the piecewise smoothness and texture properties of the diffuse component; ; These are the weight parameters.

3. The low-light image enhancement method according to claim 2, characterized in that, The step of iteratively updating the diffuse reflection component, indirect illumination component, and specular reflection component with the goal of minimizing the energy of the direct diffuse reflection-specular reflection-indirect illumination model includes: The formula for iteratively updating the diffuse reflection component is as follows: ； The formula for iteratively updating the indirect illumination component is as follows: ； The formula for iteratively updating the specular reflection component is as follows: ； In the formula: This represents the current iteration number; Let be the diffuse component after the t-th update; The next updated specular reflection components; The next update of the indirect lighting component.

4. The low-light image enhancement method according to claim 2, characterized in that, The end-to-end joint training loss function formula for the physically guided deep unfolding network is as follows: ； In the formula: To enhance the image Loss function; The structural similarity loss function; For the sparse loss function of the mirror component; The constraint loss function for the mirror inverse rendering module; These are the weighting coefficients.

5. The low-light image enhancement method according to claim 4, characterized in that, The constraint loss function of the mirror inverse rendering module includes a lighting consistency loss function and a lighting robustness loss function, as shown in the following formula: ； ； ； In the formula: These are the weighting coefficients; The specular reflection intensity coefficient is the predicted value under high light intensity conditions; The specular reflection intensity coefficient is the predicted value under low light intensity conditions; The specular index is the result of prediction under high light intensity conditions. The specular index is the result of prediction under low light intensity conditions; This is a network for estimating mirror parameters, used to predict mirror reflection parameters based on the input image and the direction of illumination; This is the low-light image at the t-th iteration; This is the predicted illumination direction vector; This is a random, small perturbation vector used to simulate the uncertainty of the illumination direction; for Learnable parameters of the network.

6. The low-light image enhancement method according to claim 1, characterized in that, The formula for the mirror inverse rendering module is as follows: ； ； In the formula: The mirror intensity coefficient; This is a surface normal diagram; The direction vector of the light; It is a half-range vector; The direction vector is fixed at the viewpoint. This refers to the highlight index.

7. The low-light image enhancement method according to claim 6, characterized in that, The specular intensity coefficient and specular index are estimated using a coarse-to-fine strategy, as shown in the following formula: ； ； ； In the formula: The initial mirror intensity coefficient; The initial highlight index; The final optimized mirror intensity coefficient; The final optimized highlight index; Reconstruct the image from the intermediate points; for Learnable parameters of the network; for Learnable parameters of the network.

8. A low-light image enhancement device based on physical illumination model constraints, characterized in that, include: The acquisition and initialization module is used to acquire the low-light image to be enhanced, and decompose the low-light image into diffuse reflection component, specular reflection component and indirect illumination component according to the direct diffuse reflection-specular reflection-indirect illumination model. The surface normal map corresponding to the low-light image is obtained using a normal estimation network; the low-light image and the surface normal map are subjected to specular inverse rendering to obtain the initial specular reflection component; the all-zero pixel matrix consistent with the low-light image is set as the initial indirect illumination component. The processing module is used to alternately update the diffuse reflection component, indirect illumination component, and specular reflection component in each iteration stage, with the goal of minimizing the energy of the direct diffuse reflection-spectral reflection-indirect illumination model: subtracting the first... The residual image obtained from the specular reflection component and the indirect illumination component of the first stage is input into the diffuse reflection update network to update the second stage. The diffuse component of the stage; subtract the first from the low-light image. The diffuse component updated in the first stage and the first The residual image obtained from the specular reflection component of the first stage is input into the indirect illumination update network to update the second stage. Indirect light component of the stage; Combined with the The diffuse reflection, indirect illumination components, and the first phase of the update are as follows: The specular reflection components of each stage are combined to form an intermediate reconstructed image. This image, combined with the surface normal map, is then processed by the specular inverse rendering module to update the next stage. The mirror reflection component of the stage; repeat the iterative update process until the energy function converges or reaches the preset termination condition; The execution module is used to superimpose the final diffuse reflection component, specular reflection component, and indirect illumination component pixel by pixel after the iteration terminates, generating and outputting the enhanced low-light image.

9. An electronic device, characterized in that, The method includes at least one processing unit and at least one storage unit, wherein the storage unit stores a computer program that, when executed by the processing unit, causes the processing unit to perform the method described in any one of claims 1 to 7.

10. A storage medium, characterized in that, It stores a computer program executable by an electronic device, which, when run on the electronic device, causes the electronic device to perform the method described in any one of claims 1 to 7.

Images