An interpretable low-light image enhancement method
By constructing an unpaired learning framework for the explicit degenerate recurrent network EDC-Net, the uninterpretability and uncontrollability of low-light image enhancement methods are solved, achieving efficient and stable image enhancement results and meeting the real-time processing requirements of nighttime monitoring and autonomous driving.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JINLING INST OF TECH
- Filing Date
- 2026-03-25
- Publication Date
- 2026-06-26
AI Technical Summary
Existing unpaired low-light image enhancement methods lack physical interpretability, are uncontrollable, have unstable training, and low inference efficiency, making it difficult to meet the real-time processing needs of nighttime monitoring and autonomous driving.
An unpaired learning framework based on the explicit degenerate recurrent network EDC-Net is constructed. The framework consists of a closed loop formed by the reinforcement branch and the degenerate branch. A three-stage progressive optimization strategy is adopted, including degenerate warm-up, master adversarial training and fine-tuning, to achieve explicit degenerate modeling and lightweight inference.
The enhancement process is interpretable and controllable, the training process converges stably, and the inference efficiency is high, significantly improving image quality and processing efficiency to meet the needs of real-time applications.
Smart Images
Figure CN122289092A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of computer vision and image processing technology, specifically to an interpretable low-light image enhancement method, applicable to scenarios requiring real-time low-light image enhancement such as night surveillance, autonomous driving, medical imaging, and digital photography. Background Technology
[0002] With the rapid development of computer vision and machine vision technologies, low-light image enhancement, as a key step in image preprocessing, plays an increasingly important role in practical scenarios such as nighttime surveillance, autonomous driving, medical imaging, and digital photography. Low-light images acquired in real-world environments generally suffer from degradation phenomena such as insufficient brightness, significant noise, and lack of contrast. Moreover, the degradation process is not simply a matter of brightness attenuation; it is often compounded by the nonlinear response and noise amplification effect of imaging devices (such as camera image signal processors, ISPs). This results in a strong coupling between brightness deviation and color error, severely affecting the visual quality of the image and significantly reducing the processing accuracy of subsequent advanced vision tasks such as object detection and image segmentation.
[0003] In recent years, deep learning-based low-light image enhancement methods have gradually become the mainstream research approach. Compared to traditional methods based on histogram equalization or Retinex theory, deep neural networks, with their powerful nonlinear modeling capabilities, can more accurately capture complex image degradation features, significantly improving the enhancement effect of low-light images. Existing deep learning-based low-light image enhancement methods are mainly divided into two categories: supervised learning and unsupervised (unpaired) learning. Supervised learning methods rely on large-scale, pixel-level precisely paired low-light-normal-light image samples for training. Although they can achieve good enhancement results, they have inherent limitations such as high sample acquisition and alignment costs and insufficient generalization ability across devices and scenes. Unpaired learning methods, on the other hand, do not rely on paired labeled data. They achieve image transformation between illumination domains through generative adversarial networks or cycle consistency constraints, effectively reducing the dependence on labeled data and becoming a research hotspot in the field of low-light image enhancement.
[0004] Currently, mainstream low-light image enhancement methods can be mainly divided into the following two categories:
[0005] (1) Low-light image enhancement methods based on supervised learning. These methods construct a large-scale, pixel-level paired low-light-normal-light image dataset and train a deep neural network to learn the mapping relationship between the two, such as RetinexNet. [1] Uformer [2]Such methods achieve good enhancement results in scenarios with sufficient paired data, but their generalization ability decreases significantly when applied to different shooting devices or unknown lighting environments. Furthermore, obtaining strictly aligned real paired samples is extremely costly, and artificially synthesized data cannot fully simulate the complex nonlinear degradation process in real-world scenes, limiting the practical application scope of these methods.
[0006] (2) Low-light image enhancement methods based on unpaired learning. These methods do not rely on paired data and achieve low-light image enhancement through generative adversarial networks or cycle consistency constraints, such as Zero-DCE. [3] RUAS [4] Such methods effectively reduce reliance on labeled data and have become a current research hotspot. However, existing unpaired methods suffer from the following technical drawbacks:
[0007] The enhancement process lacks physical interpretability and explicit constraints. Most methods use black-box networks to implicitly fit the mapping relationship from low light to normal light, ignoring the physical mechanisms such as the nonlinear response of the camera's image signal processor (ISP) and noise amplification during the imaging process. This leads to uncontrollable enhancement results and problems such as local overexposure, color shift, artifact retention, or loss of structural details.
[0008] Degradation modeling lacks parameterized representation. In cycle consistency-based training frameworks, existing methods typically use another black-box network to map the enhanced image back to the low-light domain, without explicitly modeling the degradation process in a structured and parameterized manner. This results in excessive coupling between the enhancement and degradation branches, making the training process prone to gradient oscillations and unstable convergence, leading to poor model robustness.
[0009] Inference efficiency is low. Some methods introduce complex network structures or multi-stage iterative processing in pursuit of enhanced effects, resulting in a large number of parameters and high computational overhead, making it difficult to meet the real-time processing requirements of practical application scenarios such as nighttime monitoring and autonomous driving.
[0010] Therefore, there is an urgent need for a low-light image augmentation method that is interpretable, has good augmentation effect, stable training and high inference efficiency, even without paired data. Summary of the Invention
[0011] To address the shortcomings of existing technologies, this invention provides an interpretable low-light image enhancement method, which solves the problems of uninterpretability, uncontrollability, unstable training, and low inference efficiency of existing unpaired low-light image enhancement methods, and achieves efficient and high-quality enhancement of low-light images under unpaired data.
[0012] An interpretable low-light image enhancement method is proposed, which constructs an unpaired learning framework based on the explicit degenerate recurrent network EDC-Net for low-light images. Using the input as input, low-light image enhancement is achieved through a closed loop consisting of enhancement and degradation branches. The training process employs a three-stage progressive optimization strategy, while the inference phase only runs the lightweight enhancement branch. The specific steps include:
[0013] Step 1: Convert the low-light image Input Enhancement Branch By learning pixel-by-pixel contrast adjustment factors and brightness parameters, enhanced images are generated through explicit affine mapping. ;
[0014] Step 2: Enhance the image Input Degeneracy Branch First, predict a set of physically meaningful degradation parameters. Then through an explicit degenerate operator chain Enhance the image Reconstructed as a low-light image This forms a cycle consistency constraint, and when updating degenerate branches, a gradient blocking operation is used to truncate the gradient flow of backpropagation;
[0015] Step 3: EDC-Net is trained using a three-stage progressive optimization strategy of degradation warm-up, adversarial training, and fine-tuning to ensure that the model converges smoothly under complex constraints.
[0016] Compared with the prior art, the present invention has the following significant advantages:
[0017] The enhancement process is physically interpretable and controllable: by introducing a parameterized degradation modeling mechanism into the cycle consistency framework, the enhancement branch adopts an explicit affine mapping, and the degradation branch predicts physically meaningful parameters and constructs an explicit operator chain, which solves the defects of existing black-box methods that are uninterpretable and uncontrollable, effectively avoids problems such as overexposure, color shift, artifact retention and loss of detail in the enhanced image, and improves the visual quality of the enhancement results.
[0018] Stable convergence during training: A three-stage progressive optimization strategy is adopted, which significantly reduces the coupling between the enhancement branch and the degenerate branch through degenerate warm-up, alternating update and fine-tuning in the later stage. This solves the problems of gradient oscillation and unstable convergence in unpaired closed-loop training and ensures that the model converges smoothly under complex constraints.
[0019] High inference efficiency: It adopts an asymmetric training-inference structure. During the inference phase, only the lightweight enhancement branch is run, without the need to perform the complex operation of the degenerate branch. In the experiment on the LOL-v1 dataset, the inference time is as low as 5.11ms, which is significantly better than the existing mainstream methods and can meet the real-time processing needs of real-world scenarios such as night monitoring and autonomous driving.
[0020] Excellent enhancement results: In comparative experiments on the LOL-v1 dataset, the average PSNR of this invention reached 20.731dB, SSIM reached 0.803, and LPIPS reached 0.182. All three indicators are superior to the existing mainstream supervised and unpaired low-light image enhancement methods. The enhanced image has richer details and more natural colors. Attached Figure Description
[0021] Figure 1 This is a flowchart of the overall solution of the present invention;
[0022] Figure 2 This invention enhances the internal workflow diagram of the branch;
[0023] Figure 3 This is a flowchart illustrating the internal workflow of the degenerate branch in this invention.
[0024] Figure 4 This is a comparison diagram showing the enhanced effects of the present invention and existing methods. Detailed Implementation
[0025] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The scope of protection of the present invention is not limited to the following embodiments.
[0026] The input of this invention is a low-light image. The overall scheme adopts an unpaired learning framework based on Explicit Degenerate Recurrent Network (EDC-Net), with the core structure consisting of reinforcement branches. and degenerate branches This forms a closed loop. The proposed approach for this project is as follows: First, to analyze low-light images... Input Enhancement Branch Enhanced images are generated through brightness-contrast mapping. Subsequently, degenerate branches from The algorithm predicts a set of degradation parameters with clear physical meaning, including color gain, exposure scaling, gamma correction, and blur amplitude. Here, T represents the gradient blocking operation, which truncates the backpropagating gradient flow when updating the degradation branch to prevent interference with the enhancement branch. Based on these parameters, the degradation branch is defined through an explicit degradation operator chain. Will Reconstructed as a low-light image This forms a cyclic consistency constraint. This closed-loop structure ensures that the enhancement process maintains physical interpretability and stability even without paired data. The entire training process employs a three-stage progressive optimization strategy: degradation warm-up, master adversarial training, and fine-tuning, ensuring smooth convergence of the model under complex constraints. The interpretable low-light image enhancement method of this invention is as follows: Figure 1 As shown:
[0027] 1. Enhance branches
[0028] The enhancement branch generates the enhancement result through explicit affine mapping by learning pixel-by-pixel contrast adjustment factors and luminance parameters. Given a low-light image as input... The network forward propagation yields a set of pixel-level adjustment coefficients. Six of the channels are divided into two categories of 3-channel parameter plots: . and These parameters are used to control the contrast adjustment intensity and the brightness enhancement intensity, respectively. Then, different bounded mappings are applied to the two types of control parameters to normalize the output, as shown in formula (1).
[0029]
[0030] in, Contrast control coefficient Limited to This allows the model to simultaneously express both enhanced contrast (positive value) and suppressed contrast (negative value), thereby adapting to differences in brightness and texture in different regions; The sigmoid function controls the brightness coefficient. Limited to This limits the upper bound of the bias amplitude. After obtaining... Then, a pixel-by-pixel contrast slope factor is further constructed. As shown in formula (2).
[0031]
[0032] because The angle is limited to Within range, thus avoiding exist Nearby values diverge, while preserving sufficient contrast to adjust dynamic range. Then, with... Construct a pixel-by-pixel mapping and output the enhanced image. As shown in formula (3).
[0033]
[0034] in, This indicates an element-wise multiplication operation. This indicates that each pixel will be truncated to... .
[0035] 2. Degenerate branch
[0036] Degenerate branch This is used to construct a controllable backoff process in unpaired training, thereby providing structured training information for cycle consistency. Unlike the previous black-box approach of directly fitting degenerate maps, the degenerate branch decomposes the degenerate process into two steps: parameter prediction and differentiable operator chain construction.
[0037] (1) Parameter prediction
[0038] Parameter prediction process to enhance images As input, global features are extracted through three convolutional layers, followed by adaptive average pooling and a fully connected layer to output the parameter vector. Each component is mapped to a physically bounded interval via a Sigmoid function, resulting in a degenerate set of parameters. The interval mapping of each parameter is defined as shown in formula (4).
[0039]
[0040] in, This represents the global exposure scaling parameter, used for overall brightness control. This represents the gamma correction parameter, which is used during the execution of the operator chain. and Through synergistic effect, unidirectional brightness attenuation is achieved through nonlinear mapping; This represents the three-channel color gain parameter, used to simulate differences in white balance or sensor response without excessively altering the color. This represents the Gaussian blur amplitude parameter, used to indicate mild imaging blur perturbation.
[0041] (2) Differentiable subchain
[0042] Obtaining degradation parameters Then, a fixed chain of differentiable sub-operators is constructed. This will enhance the image. Mapping back to the lower optical domain to form a closed loop: The order of the operator chain is: color gain Exposure scaling Gamma correction Fuzziness amplitude. The specific forms of each operator are as follows:
[0043] a) Color gain and exposure scaling: ,in, This represents element-wise multiplication, using constraints. This ensures that the degradation process proceeds along the direction of monotonic brightness decay.
[0044] b) Gamma correction: ,in, , guarantee Within the pixel domain, this transformation results in an overall darkening state, which helps to revert the image from the enhancement domain to the low-light domain distribution.
[0045] c) Blur amplitude: ,in, This represents the convolution operation. The kernel size varies with... Adaptive adjustment, convolution uses To maintain the same resolution. Final output. Cut off to the effective range This is to prevent pixels from going out of bounds.
[0046] 3. Three-stage incremental optimization strategy
[0047] In unpaired closed-loop training, simultaneous updates of the augmentation and degenerate branches can easily lead to gradient oscillations. To reduce the coupling between the two branches and improve convergence stability, EDC-Net employs a three-stage progressive optimization strategy, including degenerate warm-up, adversarial training, and fine-tuning.
[0048] (1) Degradation preheating stage
[0049] The enhancement branch is frozen, and only the degradation branch is optimized, allowing the degradation branch to learn a backtracking mapping from the enhancement domain to the low-light domain within the explicit physical parameter space. The purpose of this stage is to provide a relatively stable degradation baseline and avoid gradient divergence.
[0050] (2) Main confrontation training phase
[0051] During the main adversarial training phase, the two branches are updated alternately rather than simultaneously. Specifically, in each iteration, the degenerate branch is updated first, followed by the enhancement branch. When updating the degenerate branch, gradient blocking is applied to the enhancement output, preventing the gradients of the degenerate branch from being backpropagated to the enhancement branch, thus avoiding interference with the enhancement parameters during the degenerate learning phase. When updating the enhancement branch, the parameters of the degenerate branch are frozen, but its differentiable degenerate path is retained for backpropagation, allowing the enhancement branch to learn a stable enhancement mapping under fixed explicit degenerate constraints. This alternating mechanism controls the gradient flow direction, preventing the two branches from mutually restricting each other and causing convergence instability.
[0052] (3) Fine-tuning stage
[0053] After primary adversarial training, the degenerate branch is frozen, and the augmentation branch is optimized only with a smaller learning rate, with the loss weights adjusted accordingly. This stage no longer alters the degenerate mapping but further improves texture detail and color consistency under stable constraints.
[0054] This invention employs an explicit degradation recurrent network composed of enhancement and degradation branches. The enhancement branch is responsible for the adaptive mapping from low-light images to normal-light images, while the degradation branch is responsible for predicting physically meaningful degradation parameters and implementing closed-loop constraints through a chain of explicit degradation operators. Furthermore, it utilizes a three-stage progressive optimization strategy and executes only the lightweight enhancement branch during the inference phase. These are the core technical solutions of this invention, providing a technical foundation for interpretable and efficient enhancement of low-light images under unpaired conditions.
[0055] This invention discloses an interpretable low-light image enhancement method based on an explicit degradation recurrent network. Compared with the mainstream low-light image enhancement methods described in Part II, this invention solves the defects of existing methods, such as uninterpretable enhancement process and lack of physical constraints, by constructing a closed-loop learning framework composed of enhancement branches and degradation branches. It avoids problems such as overexposure, artifacts, or loss of detail in the enhanced image, and significantly improves the controllability and visual quality of the enhancement results. As shown in Figure 4, the image enhanced by the method of this invention has richer details, more natural colors, and no artifacts or overexposure.
[0056] Meanwhile, this invention employs a three-stage progressive optimization strategy (degradation warm-up, adversarial training, and fine-tuning), effectively reducing the coupling between the enhancement and degradation branches and resolving the issues of gradient oscillation and convergence instability during training, enabling the model to converge smoothly under complex constraints. Furthermore, this invention adopts a training-inference structure, where only the lightweight enhancement branch is run during the inference phase, significantly improving processing efficiency while ensuring enhancement effectiveness and overcoming the shortcomings of existing methods in meeting real-time application requirements. Table 1 shows the comparison results of the average PSNR, SSIM, and LPIPS values of the method of this invention with existing mainstream low-light image enhancement methods on the LOL-v1 dataset under the same experimental environment. The method of this invention achieves the best results in all three metrics.
[0057] All comparative experiments were run under the same conditions: Intel(R) Core(TM) i7-14650HX CPU, 16 GB of memory, Windows 11 operating system, and NVIDIA RTX 5060 GPU. Table 2 shows the comparison results of the inference time of the proposed method and existing mainstream low-light image enhancement methods on the LOL-v1 dataset under the same experimental conditions. The proposed EDC-Net method has the best inference time and more outstanding real-time processing efficiency.
[0058] method PSNR (dB) ↑ SSIM↑ LPIPS↓ RetinexNet 16.774 0.462 0.474 Uformer 18.547 0.730 0.321 Zero-DCE 14.857 0.589 0.335 RUAS 16.398 0.537 0.350 EDC-Net 20.731 0.803 0.182
[0059] Table 1 Comparison results between the present invention and previous methods
[0060] method Time (ms) RetinexNet 42.8 Uformer 156 Zero-DCE 5.3 RUAS 18.6 EDC-Net 5.11
[0061] Table 2 Comparison of inference time between the present invention and previous methods
[0062] This invention provides an interpretable low-light image enhancement method, which is not limited to the specific implementations of the enhancement branch, degradation branch, and three-stage progressive optimization strategy described in the specification and embodiments. Any equivalent changes or modifications made based on the structure, features, and principles of this invention's technical solution, such as using different neural network architectures to decouple enhancement and degradation, introducing color temperature adjustment, lens distortion simulation, or other physically meaningful combinations of degradation parameters, changing the number of progressive training stages or the form of the loss function, or performing lightweight processing such as quantization and pruning on the enhancement branch during the inference stage, as long as the core idea remains the same—constructing closed-loop constraints through explicit parameterized degradation modeling to achieve interpretable low-light image enhancement—should be included within the scope of this patent application.
Claims
1. An interpretable low-light image enhancement method, characterized in that, Construct an unpaired learning framework based on the explicit degenerate recurrent network EDC-Net for low-light images. Using the input as input, low-light image enhancement is achieved through a closed loop consisting of enhancement and degradation branches. The training process employs a three-stage progressive optimization strategy, while the inference phase only runs the lightweight enhancement branch. The specific steps include: Step 1: Convert the low-light image Input Enhancement Branch By learning pixel-by-pixel contrast adjustment factors and brightness parameters, enhanced images are generated through explicit affine mapping. ; Step 2: Enhance the image Input Degeneracy Branch First, predict a set of physically meaningful degradation parameters. Then through an explicit degenerate operator chain Enhance the image Reconstructed as a low-light image This forms a cycle consistency constraint, and when updating degenerate branches, a gradient blocking operation is used to truncate the gradient flow of backpropagation; Step 3: EDC-Net is trained using a three-stage progressive optimization strategy of degradation warm-up, adversarial training, and fine-tuning to ensure that the model converges smoothly under complex constraints.
2. The interpretable low-light image enhancement method according to claim 1, characterized in that, The enhanced branch generated in step 1 generates an enhanced image. The specific process is as follows: Low-light images Pixel-level adjustment coefficients are obtained through network forward propagation. ,Will Raw contrast parameters divided into 3 channels and brightness raw parameters ; right and By applying bounded mappings, contrast control coefficients are obtained. Brightness control coefficient ,in It is the hyperbolic tangent function. For the Sigmoid function; Based on contrast control coefficient Constructing the pixel-by-pixel contrast slope factor ; by Constructing a pixel-by-pixel mapping using the formula Generate enhanced images ,in This is an element-wise multiplication operation. This is a pixel truncation operation.
3. The interpretable low-light image enhancement method according to claim 1, characterized in that, The degradation parameters mentioned in step 2 ,in Global exposure scaling parameters, For gamma correction parameters, For three-channel color gain parameters, The Gaussian blur amplitude parameter; Degeneracy branch predicts degradation parameters The specific process is as follows: to enhance the image As input, global features are extracted through three convolutional layers, and the output parameter vector is obtained through adaptive average pooling and fully connected layers. Each component is mapped to a physically bounded interval via a Sigmoid function to obtain the degradation parameters. The mapping relationship is as follows: ； ； ； 。 4. The interpretable low-light image enhancement method according to claim 3, characterized in that, The explicit degenerate operator chain described in step 2 The process is executed in the following order: color gain → exposure scaling → gamma correction → blur amplitude. Color gain and exposure scaling: ,in, This represents element-wise multiplication, using constraints. This ensures that the degradation process proceeds along the direction of monotonic brightness decay. Gamma correction: ,in, , guarantee Within the pixel domain, this transformation results in an overall darkening state, which helps to revert the image from the enhancement domain to the low-light domain distribution. Blur range: ,in, This represents the convolution operation. The kernel size varies with... Adaptive adjustment, convolution uses To maintain the same resolution; Final output Cut off to the effective range This is done to avoid pixels going out of bounds, resulting in a reconstructed low-light image.
5. The interpretable low-light image enhancement method according to claim 1, characterized in that, The specific process of the three-stage incremental optimization strategy described in step 3 is as follows: Degradation warm-up stage: Freeze the enhancement branch and optimize only the degradation branch, so that the degradation branch learns the back-down mapping from the enhancement domain to the low-light domain in the explicit physical parameter space; Main adversarial training phase: The degenerate branch and the augmentation branch are updated alternately. First, the degenerate branch is updated and gradient blocking is applied to the augmentation output. Then, the augmentation branch is updated and the parameters of the degenerate branch are frozen. The differentiability of the degenerate link is preserved to participate in backpropagation. Fine-tuning phase: Freeze the degenerate branch, continue to optimize the enhancement branch with a smaller learning rate, and adjust the loss weights to improve image texture details and color consistency under stable constraints.
6. The interpretable low-light image enhancement method according to any one of claims 1-5, characterized in that, In the closed loop formed by the enhancement branch and the degradation branch, exposure loss... and cycle consistency loss Apply loss constraints to the model.
7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the interpretable low-light image enhancement method according to any one of claims 1-6.
8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the interpretable low-light image enhancement method according to any one of claims 1-6.