A non-local and plug-and-play based low-light image enhancement method and system
By introducing auxiliary variables and guided filters, the image enhancement method based on the Retinex model solves the problems of unnatural image enhancement and difficulty in obtaining datasets in existing technologies. It achieves efficient image enhancement under low light conditions, preserves edges and details, and reduces artifacts and noise.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-24
- Publication Date
- 2026-03-27
AI Technical Summary
Existing Retinex-based image enhancement techniques result in unnatural image enhancement effects, artifacts, and noise amplification when the illumination component is ignored. Furthermore, large training datasets are difficult to obtain, leading to unsatisfactory visible light artifacts when features are mismatched.
A nonlocal and plug-and-play low-light image enhancement method is adopted. The image enhancement model is constructed through the Retinex model, auxiliary variables are introduced and transformed into Lagrangian functions, and the sub-optimization function is solved using the alternating direction multiplier method. Edge-preserving filtering is performed in combination with a guided filter to optimize the image enhancement results.
Under low-light conditions, it maintains the smoothness of the illumination layer and preserves prominent edges, effectively distinguishes structural elements and texture areas, prevents noise effects, improves image enhancement, reduces artifacts and noise, and maintains the naturalness of the image.
Smart Images
Figure CN120782644B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of image processing, and in particular to a non-local and plug-and-play based image enhancement method and system. BACKGROUND
[0002] The Retinex-based method constructs an input image by multiplying the illumination component and the reflection component at each pixel. The Retinex concept explains how the human visual system (HVS) perceives color and light, mainly focusing on the reflection component and taking it as the improved output image.
[0003] However, the existing Retinex-based image enhancement technology still has the following defects:
[0004] 1. When the illumination component is ignored, the image enhancement result tends to be over-enhanced and unnatural, resulting in poor image enhancement effect, and may also appear artifacts, noise amplification, loss of details, etc.
[0005] 2. In most current learning-based methods, the performance of image enhancement still heavily depends on large training data sets, which are not easy to obtain when collected on a large scale. In addition, when the training data set and the test image have different characteristics, the result usually appears undesirable visible light artifacts. SUMMARY
[0006] The purpose of the present application is to provide a non-local and plug-and-play based low-light image enhancement method and system, which solves the problems of poor image enhancement effect, high difficulty in obtaining data sets, and undesirable visible light artifacts in the prior art.
[0007] To achieve the above purpose, the present application provides a non-local and plug-and-play based low-light image enhancement method, comprising the following steps:
[0008] S1, pre-processing the parameters of the input image, constructing an image enhancement model based on the Retinex model, introducing an auxiliary variable in the image enhancement model, and converting the image enhancement model to obtain an image enhancement Lagrangian function;
[0009] S2, constructing a sub-optimization function based on the image enhancement Lagrangian function, and solving the sub-optimization function by using the alternating direction multiplier method (ADMM) to obtain an optimization update value;
[0010] S3, performing edge-preserving filtering on the optimization update value using a guided filter to obtain an optimization filtering result;
[0011] S4, updating the image enhancement model based on the optimization update value and the optimization filtering result to obtain an image enhancement result.
[0012] In some embodiments of the present application, in S1, the parameters of the input image are preprocessed, and an image enhancement model is constructed based on a Retinex model, which includes:
[0013] S11, normalize the pixel value of the input image to the range of [0, 1], and convert the input image to the hue-saturation-brightness color space;
[0014] S12, construct an image enhancement model based on the Retinex model, and perform brightness enhancement on the brightness channel, and the expression of the image enhancement model is:
[0015]
[0016] wherein T is the illumination layer, R is the updated reflectance layer, is an operator for multiplying each corresponding element of the two layers, T0 is an initial illumination parameter, q and F are auxiliary parameters, S is the input image, f(T) is an illumination function, w1 and w2, and δ are importance weight parameters, α and β are fractional order parameters, and the value range is (1, 2), W R is a weight matrix, and are fractional order derivative filters.
[0017] In some embodiments of the present application, in S1, auxiliary variables are introduced into the image enhancement model, and the image enhancement model is transformed to obtain an image enhancement Lagrangian function, which includes:
[0018] S13, two auxiliary variables are introduced into the image enhancement model to obtain an augmented Lagrangian function of the image enhancement model, and the expression is:
[0019]
[0020] wherein L and f(I) are auxiliary variables, f(I)=f(T), μ is a positive penalty factor, Z and Y are both Lagrange multipliers, and <·,·> is a matrix inner product operator, is an image enhancement Lagrangian function.
[0021] In some embodiments of the present application, in S2, the sub-optimization function is constructed based on the image enhancement Lagrangian function, which includes:
[0022] S21, remove the term irrelevant to the updated reflectance layer R in the image enhancement Lagrangian function to obtain an R sub-optimization function, and the expression is:
[0023]
[0024] wherein k is the order of derivation, r, t, and s are the vector forms of R, T, and S respectively, and T and W respectively R The diagonal matrix form, where the matrix is N×N dimensional, A d Let R be the derivative of R, N be the total number of pixels, and D be the derivative of R. h and D v All are Toplitz matrices with backward difference discrete derivative operators;
[0025] S22. Extract the term related to the illumination layer T from the image enhancement Lagrangian function to obtain the T sub-optimization function, the expression of which is:
[0026]
[0027] in, Let Y be an identity matrix of a specific size, and y be the derivative of Y.
[0028] In some embodiments of this application, S2 further includes:
[0029] S23. Extract the terms related to the auxiliary variable L from the image enhancement Lagrangian function to obtain the L sub-optimization function, the expression of which is:
[0030]
[0031] S24. Extract the terms related to the auxiliary variable I from the image enhancement Lagrangian function to obtain the I sub-optimization function, the expression of which is:
[0032]
[0033] The sub-optimization function I is converted into a guided filter, expressed as:
[0034]
[0035] in, It is a parametric Toplitz matrix.
[0036] In some embodiments of this application, in step S3, a guided filter is used to perform edge-preserving filtering on the optimized update value, and the optimized filtering result includes:
[0037] S31. The surface of the object in the illumination layer is smoothed using a guided filter, while sharp edges are preserved. The expression for the guided filter in image processing is:
[0038]
[0039] Where I1 and I2 are both guided filters, t g1 t g2 φ1 and φ2 are both filter parameters;
[0040] S32, combine I1 and I2 in the above formula by threshold to obtain an expression as follows:
[0041]
[0042] wherein I3 is a guided filter of I1 and I2, thres is a gamma correction value of a preset intensity, and slope is a slope parameter;
[0043] S33, based on the result of step S32, processing the input image, applying I2 processing to the pixels of the input image below the threshold intensity, and using I1 processing to the pixels above the threshold intensity, to obtain an optimized filter result after filter processing.
[0044] In some embodiments of the present application, a Retinex-based low-light image enhancement system is also disclosed, comprising:
[0045] (1) a model construction module: for pre-processing the parameters of the input image, constructing an image enhancement model based on the Retinex model, introducing an auxiliary variable in the image enhancement model, and converting the image enhancement model to obtain an image enhancement Lagrangian function;
[0046] (2) an optimization update module: for constructing a sub-optimization function based on the image enhancement Lagrangian function, and solving the sub-optimization function by using the alternating direction multiplier method to obtain an optimization update value;
[0047] (3) an optimized filter module: for performing edge-preserving filtering on the optimization update value using a guided filter to obtain an optimized filter result;
[0048] (4) an image enhancement module: for updating the image enhancement model based on the optimization update value and the optimized filter result to obtain an image enhancement result.
[0049] The advantages and beneficial effects of the present application relative to the prior art are:
[0050] 1. The present application considers the application of low-light enhancement under the characteristics of illumination and reflection. The optimal illumination layer should maintain smoothness in a segmented manner, while only retaining the most prominent edges.
[0051] 2. Considering that a single detector may not work efficiently due to significant noise, the present application adds local fluctuations and fractional derivatives to accurately distinguish between structural elements, textures, and noise regions.
[0052] The technical solutions of the present application will be further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF DRAWINGS
[0053] Figure 1A flowchart of the steps of a non-local and plug-and-play based low-light image enhancement method according to an embodiment of the present application;
[0054] Figure 2 A framework of a low-light image enhancement system based on Retinex according to an embodiment of the present application;
[0055] Figure 3 A comparison of the results of low-light image enhancement by the present application and existing Retinex variational methods, where (a) is the result of enhancement by the RRM method, (b) is the result of enhancement by the HVM method, (c) is the result of enhancement by the Hao et al. method, and (d) is the result of enhancement by the present application;
[0056] Figure 4 The influence of the guided filter and Lq norm introduced by the present application on the result of enhancement, where (a) is a low-light RGB image, (b) is the result of enhancement using only the Lq norm, (c) is the result of enhancement using the guided filter F1 combined with the Lq norm, (d) is the result of enhancement using the guided filter F2 combined with the Lq norm, and (e) is the result of enhancement using the hybrid guided filter F3 combined with the Lq norm;
[0057] Figure 5 A comparison of the processing time of the guided filter introduced by the present application and several existing filter-based enhancement methods, where (a) is a low-light RGB image, (b) is the result of enhancement using the L0 gradient minimization method (19.5 seconds), (c) is the result of enhancement using the relative gradient filter method (26.5 seconds), (d) is the result of enhancement using the weighted least squares method (6.1 seconds), and (e) is the result of enhancement using the guided filter proposed in this paper (5.3 seconds);
[0058] Figure 6 A comparison of the results of low-light RGB image enhancement by the present application and existing deep learning-based methods, where (a) is a low-light RGB image, (b) is the result of enhancement by the RUAS method, (c) is the result of enhancement by the Zero-DCE method, (d) is the result of enhancement by the URetinex-Net method, and (e) is the result of enhancement by the present application;
[0059] Figure 7 A flowchart of the method according to an embodiment of the present application. DETAILED DESCRIPTION
[0060] In the description of the present application, it should be noted that the terms "upper", "lower", "inner", "outer" and the like indicate the orientation or positional relationship shown in the drawings, or the orientation or positional relationship in which the product of the present application is usually placed, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application. In the description of the present application, it should be noted that, unless otherwise explicitly specified and limited, the terms "arrangement", "installation", "connection" should be understood broadly, for example, it can be fixedly connected, or it can be detachably connected, or integrally connected; it can be mechanically connected, or it can be electrically connected; it can be directly connected, or it can be indirectly connected through an intermediate medium; it can be the communication between the two elements inside. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.
[0061] The embodiments of the present application will be described in detail below with reference to the drawings.
[0062] As Figure 1 shown, the present application provides a non-local and plug-and-play low-light image enhancement method, comprising the following steps:
[0063] S1, pre-processing the parameters of the input image, constructing an image enhancement model based on the Retinex model, introducing an auxiliary variable in the image enhancement model, and converting the image enhancement model to obtain an image enhancement Lagrange function;
[0064] S2, constructing a sub-optimization function based on the image enhancement Lagrange function, and solving the sub-optimization function by using the alternating direction multiplier method ADMM to obtain an optimization update value;
[0065] S3, using a guided filter to perform edge-preserving filtering on the optimization update value to obtain an optimization filtering result;
[0066] S4, updating the image enhancement model based on the optimization update value and the optimization filtering result to obtain an image enhancement result.
[0067] Specifically, the classical Retinex model is based on physical principles, and the observed image S is explained as the product of the illumination layer T and the reflectivity layer R. According to the Retinex theory, the illumination layer T corresponds to the illumination intensity of the object, and is expected to be piecewise continuous without any texture. The reflectivity layer R represents the inherent characteristics of the captured object, which contains details and textures in the observed image.
[0068] Starting from an RGB image, the present application first normalizes the pixel values to the range [0, 1], and then converts the image to the Hue-Saturation-Value (HSV) color space. We apply an enhancement technique specifically to the V (Value) channel. The input image S is enhanced on the V channel to obtain S * Finally, the modified V channel is combined with the original H and S channels, and converted back to the RGB color space to generate the final output. The present application introduces a new Retinex framework to solve the energy minimization problem.
[0069] In some embodiments of the present application, in S1, the parameters of the input image are preprocessed, and the image enhancement model is constructed based on the Retinex model, comprising:
[0070] S11, normalizing the pixel values of the input image to the range [0, 1], and converting the input image to the Hue-Saturation-Value color space;
[0071] S12, constructing an image enhancement model based on the Retinex model, and performing brightness enhancement on the brightness channel, the expression of the image enhancement model is:
[0072]
[0073] wherein T is the illumination layer, R is the updated reflectance layer, is an operator for multiplying each corresponding element of the two layers, T0 is the initial illumination parameter, q, F are auxiliary parameters, S is the input image, f(T) is the illumination function, w1, w2, and δ are importance weight parameters, α and β are fractional order parameters, the value range is (1, 2), W R is a weight matrix, are fractional order derivative filters.
[0074] In some embodiments of the present application, in S1, auxiliary variables are introduced into the image enhancement model, and the image enhancement model is transformed to obtain an image enhancement Lagrangian function, comprising:
[0075] S13, introducing two auxiliary variables into the image enhancement model to obtain an augmented Lagrangian function of the image enhancement model, the expression is:
[0076]
[0077] wherein L and f(I) are auxiliary variables, f(I)=f(T), μ is a positive penalty factor, Z and Y are both Lagrange multipliers, <·,·> is a matrix inner product operator, is the image enhancement Lagrangian function.
[0078] Specifically, the above optimization problem can be solved using the traditional ADMM method. In this process, we iteratively update each variable while keeping other variables fixed at the values of the previous iteration.
[0079] In some embodiments of the present application, in S2, constructing a sub-optimization function based on the image enhancement Lagrangian function comprises:
[0080] S21, removing the term irrelevant to the updated reflectance layer R in the image enhancement Lagrangian function, obtaining the R sub-optimization function, the expression is:
[0081]
[0082] Wherein, k is the derivation order, r, t, s are the vector forms of R, T and S respectively, And are the diagonal matrix forms of T and W R , the matrix is N x N dimension, A d is the derivative function of R, N is the total number of pixels, D h and D v are both Toeplitz matrices with backward difference discrete derivative operator;
[0083] It is to be understood that in order to avoid the complexity of direct matrix inversion caused by large dimension of matrix, we use the preconditioned conjugate gradient method to solve the above equation.
[0084] S22, extracting the term related to the illumination layer T in the image enhancement Lagrangian function, obtaining the T sub-optimization function, the expression is:
[0085]
[0086] Wherein, is a unit matrix of a certain size, and y is the derivative form of Y.
[0087] It is to be understood that l and z can be processed in the horizontal and vertical directions respectively.
[0088] In some embodiments of the present application, in S2, it further comprises:
[0089] S23, extracting the term related to the auxiliary variable L in the image enhancement Lagrangian function, obtaining the L sub-optimization function, the expression is:
[0090]
[0091] It is to be understood that since the closed-form solution of the above L q approximate mapping does not exist, we must use L q shrinkage mapping instead.
[0092] S24, extract the term related to auxiliary variable I in the image enhancement Lagrange function, get I sub-optimization function, expression is:
[0093]
[0094] Regarding the above formula f(I), we do not need to define it and solve its approximate mapping, because the sub-optimization function can be explained as a filtering operation within the PnPADMM framework. In order to obtain the ideal illumination mapping, we will replace the image denoiser in the initial PnPADMM with an edge-preserving filter (a steering filter), and convert the I sub-optimization function into a steering filter, expression is:
[0095]
[0096] Where, is a parameterized tophat matrix.
[0097] In some embodiments of the present application, in S3, the edge-preserving filter is used to perform edge-preserving filtering on the optimization update value to obtain an optimization filtering result, including:
[0098] It should be understood that we choose the steering filter as the edge-preserving filter. The steering filter, as an edge-preserving smoothing operator, provides high efficiency and fast performance. Compared with other filters, it shows relatively low computational complexity.
[0099] The illumination layer utilizes the advantageous characteristics of the steering filter to smooth the object surface while effectively preserving sharp boundaries, benefiting from the use of L q norm.
[0100] S31, utilize the steering filter to perform smoothing processing on the object surface of the illumination layer while preserving sharp boundaries, expression of the steering filter performing image processing is:
[0101]
[0102]
[0103] Wherein, I1 and I2 are both steering filters, t g1 , g2 , are filtering parameters;
[0104] Specifically, both filters are self-steering, and the parameter is set to t g1 =t g2 =5, and In the case of I1, although the low intensity region is significantly smoothed, the boundary between the sky and the house is still clearly visible, and in the logarithmic domain, the low intensity region is amplified, therefore, I2 retains more obvious dark structures and enhances the region with lower brightness. However, a significant halo effect appears around the house and the trees. Based on this, the present application considers using a threshold to combine the advantages of the two results. The filter finally applied to I is I3.
[0105] S32, threshold combination of I1 and I2 in the above formula is obtained as follows:
[0106]
[0107] Wherein, I3 is a guided filter combining I1 and I2, thres is a preset intensity gamma correction value, and slope is a slope parameter.
[0108] S33, based on the result of step S32, the input image is processed, I2 processing is applied to the pixels of the input image below the threshold intensity, and I1 processing is used for the pixels above the threshold intensity, to obtain an optimized filter result after filter processing.
[0109] In some embodiments of the present application, as shown in Figure 2 It is also disclosed that a low-light image enhancement system based on Retinex comprises:
[0110] A model construction module is configured to pre-process parameters of an input image, construct an image enhancement model based on a Retinex model, introduce an auxiliary variable into the image enhancement model, and convert the image enhancement model to obtain an image enhancement Lagrange function.
[0111] An optimization update module is configured to construct a sub-optimization function based on the image enhancement Lagrange function, and solve the sub-optimization function by using an alternating direction multiplier method to obtain an optimization update value.
[0112] An optimization filter module is configured to perform edge-preserving filtering on the optimization update value by using a guided filter to obtain an optimized filter result.
[0113] An image enhancement module is configured to update the image enhancement model based on the optimization update value and the optimized filter result to obtain an image enhancement result.
[0114] The present application has the following advantages and beneficial effects compared with the prior art:
[0115] 1. The present application considers the application of low-light enhancement under the illumination and reflection characteristics, and the optimal illumination layer should maintain smoothness in a segmented manner, while only retaining the most prominent edges.
[0116] 2. Considering that a single detector can not work efficiently due to significant noise, we incorporate local fluctuation and fractional derivative in our consideration to accurately distinguish between structure elements, texture and noise regions.
[0117] The embodiments of the present application are explained in detail below with specific examples. Several experimental results are shown here to highlight the effectiveness of the proposed method from both qualitative and quantitative perspectives. The experiments are conducted using MATLAB R2014a on a system equipped with 64 GB RAM and a 2.5 GHz processor. Throughout the experiments, the empirical values of the parameters are assigned the following values: w1 = 0.3, w2 = 0.3, δ = 0.01, p = 1.5, q = 0.4, a = 1.4, and β = 1.8. If necessary, the preconditioned conjugate gradient (PCG) method is used to perform matrix inversion using the pcg function of MATLAB.
[0118] First, data collection is performed. In order to comprehensively evaluate the proposed method, we test our method on images from different scenes. These test images are collected from two different datasets. The first test dataset contains 100 images from multiple public sources: LIME
[12] , DICM
[24] , VV
[37] , NASA
[38] , CC
[40] and MEF
[39] . Given that the MEF dataset contains multi-exposure images, only low-light images in each exposure group are selected for evaluation. The second experimental dataset contains 500 processed images, numbered #4501-#5000, from the Massachusetts Institute of Technology (MIT). This study uses the Adobe FiveK dataset. Images #0001 to #4500 are omitted as they are usually used as training data in learning-based methods. Due to the memory usage restrictions imposed by previous studies, we scale the images in the CC and Massachusetts Institute of Technology (MIT) Adobe FiveK datasets to 65 times their original size.
[0119] We evaluate the performance of various edge-preserving filters in generating high-fidelity illumination maps. These filter-based methods include L0 gradient minimization (L0-GM), rolling guidance filter (RGF), weighted least squares filter (WLS), and the proposed guidance filter. L0-GM combines unrelated local regions into a single region, often resulting in noticeable artifacts. RGF and WLS cannot effectively smooth structures, leading to reduced contrast and reduced visibility in the enhanced images. In addition, the computational cost of L0-GM and RGF is also high. Compared with other techniques, the proposed guidance filter shows better performance and faster processing speed. Considering both visual quality and processing efficiency, we choose the self-guided filter as the proposed model.
[0120] In challenging low-light scenes, the proposed method preserves fine details and ensures natural exposure, while other low-light image enhancement techniques struggle to accurately reconstruct buildings and rocks. LIME over-amplifies high-intensity regions, while RRM, Hao et al., and LR3M oversmooth fine textures. The enhanced results produced by the LM3R method are also accompanied by noticeable artifacts. The reader can observe the edges of the building amplification part. The results produced by the WLS method show color distortion. For more details, the reader can observe the sky region. EFF enhances local details in a conservative manner, and the flowchart of the method of the present application is shown in Figure 7
[0121] As shown in Figure 3 , the results of low-light image enhancement by the present application and existing Retinex variational methods are compared, where (a) is the enhanced result of the RRM method, (b) is the enhanced result of the HVM method, (c) is the enhanced result of the Hao et al. method, and (d) is the enhanced result of the present application;
[0122] As shown in Figure 4 , the effects of the guided filter and Lq norm introduced by the present application on the enhanced results are shown, where (a) is a low-light RGB image, (b) is the enhanced result using only the Lq norm, (c) is the enhanced result using the guided filter F1 combined with the Lq norm, (d) is the enhanced result using the guided filter F2 combined with the Lq norm, and (e) is the enhanced result using the hybrid guided filter F3 combined with the Lq norm;
[0123] As shown in Figure 5 , a comparison of the processing time of the guided filter introduced by the present application and several existing filter-based enhancement methods is shown, where (a) is a low-light RGB image, (b) is the use of the L0 gradient minimization method (19.5 seconds), (c) is the use of the relative gradient filter method (26.5 seconds), (d) is the use of the weighted least squares method (6.1 seconds), and (e) is the guided filter proposed in this paper (5.3 seconds);
[0124] As shown in Figure 6 , a comparison of the results of low-light RGB image enhancement by the present application and existing deep learning-based methods is shown, where (a) is a low-light RGB image, (b) is the enhanced result of the RUAS method, (c) is the enhanced result of the Zero-DCE method, (d) is the enhanced result of the URetinex-Net method, and (e) is the enhanced result of the present application.
[0125] In summary, the method of the present application successfully improves the visibility of dark areas while preserving additional textures and finer details, in addition, the present application also prevents over-enhancement of bright areas and maintains color integrity.
[0126] In the present application, unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. If there is a conflict between the definitions in the specification and those in the patent specification, the definitions in the specification are intended to prevail. In addition, the terms used herein are merely for the purpose of describing the embodiments of the present application and are not intended to limit the present application.
[0127] Finally, it should be noted that the above examples are only used to illustrate the technical solutions of the present application, but not to limit the present application. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can still be modified or equivalently replaced, and these modifications or equivalent replacements should not make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present application.
Claims
1. A low-light image enhancement method based on nonlocal and plug-and-play principles, characterized in that, Includes the following steps: S1. Preprocess the parameters of the input image, construct an image enhancement model based on the Retinex model, introduce auxiliary variables into the image enhancement model, and transform the image enhancement model into the image enhancement Lagrangian function; In step S1, the parameters of the input image are preprocessed, and an image enhancement model is constructed based on the Retinex model, including: S11. Normalize the pixel values of the input image to the range of [0, 1], and convert the input image to the hue-saturation-brightness color space; S12. Construct an image enhancement model based on the Retinex model, and enhance the brightness of the luminance channel. The expression of the image enhancement model is: Where T is the illumination layer and R is the refresh reflectivity layer. An operator that performs multiplication on each corresponding element in two layers. Here are the initial lighting parameters, and q and F are auxiliary parameters. For the input image, For the illumination function, , , and These are all importance weight parameters. and This is a fractional parameter, with a value range of (1, 2). This is the weight matrix. , All are fractional derivative filters; In step S1, introducing auxiliary variables into the image enhancement model and transforming the image enhancement model into an image enhancement Lagrangian function includes: S13. Introducing two auxiliary variables into the image enhancement model yields the augmented Lagrangian function of the image enhancement model, expressed as: in, and As an auxiliary variable, , As a positive penalty factor, and All are Lagrange multipliers. This is the matrix inner product operator. Enhance the Lagrangian function for the image; S2. Construct a sub-optimization function based on the image enhancement Lagrangian function, and solve the sub-optimization function using the alternating direction multiplier method to obtain the optimized update value; S3. Use a guided filter to perform edge-preserving filtering on the optimized update value to obtain the optimized filtering result; S4. Update the image enhancement model based on the optimized update value and the optimized filtering result to obtain the image enhancement result.
2. The low-light image enhancement method based on nonlocal and plug-and-play as described in claim 1, characterized in that, In step S2, constructing the sub-optimization function based on the image enhancement Lagrangian function includes: S21. Remove the Lagrangian function from the image enhancement layer and update the reflectivity layer. Irrelevant items, obtained The sub-optimization function is expressed as: ； ； in, To find the order of the derivative, 、 、 They are respectively 、 、 vector form, and They are respectively and The diagonal matrix form, the matrix is dimension, for R The derivative function, Total number of pixels and All are Toplitz matrices with backward difference discrete derivative operators; S22. Extract the image enhancement Lagrangian function from the illumination layer. T Related items, obtained T The sub-optimization function is expressed as: ； in, For an identity matrix of a specific size, for The derivative form of .
3. The low-light image enhancement method based on nonlocal and plug-and-play according to claim 2, characterized in that, S2 also includes: S23. Extracting auxiliary variables from the image enhancement Lagrange function. L Related items, obtained L The sub-optimization function is expressed as: ； S24. Extracting auxiliary variables from the image enhancement Lagrangian function. I Related items, obtained I The sub-optimization function is expressed as: ； Will I The sub-optimization function is converted into a guided filter, expressed as: ； in, It is a parametric Toplitz matrix.
4. The low-light image enhancement method based on nonlocal and plug-and-play according to claim 3, characterized in that, In step S3, a guided filter is used to perform edge-preserving filtering on the optimized update value, and the optimized filtering results include: S31. The surface of the object in the illumination layer is smoothed using a guided filter, while sharp edges are preserved. The expression for the guided filter in image processing is: ； in, and All are guided filters. All of these are filter parameters; S32, in the above formula and Thresholding yields the following expression: in, For comprehensive and The guided filter, The gamma correction value is the preset intensity. These are the slope parameters; S33, based on the result of step S32, process the input image, and apply [method / applied] to pixels in the input image with intensities below a threshold. Processing is applied to pixels with intensity exceeding a threshold. The process is repeated to obtain the optimized filtering result.
5. A Retinex-based low-light image enhancement system, performing the non-local and plug-and-play low-light image enhancement method as described in any one of claims 1-4, characterized in that, include: (1) Model building module: used to preprocess the parameters of the input image, build an image enhancement model based on the Retinex model, introduce auxiliary variables into the image enhancement model, and transform the image enhancement model into the image enhancement Lagrangian function; (2) Optimization and update module: used to construct sub-optimization functions based on image enhancement Lagrangian function, and solve the sub-optimization functions using alternating direction multiplier method to obtain the optimized update value; (3) Optimization filtering module: used to perform edge-preserving filtering on the optimized update value using a guided filter to obtain the optimization filtering result; (4) Image enhancement module: used to update the image enhancement model based on the optimized update value and the optimized filtering result to obtain the image enhancement result.