Low-illumination image enhancement method

Abstract

The invention relates to the field of image enhancement, in particular to a low-light image enhancement method, which comprises the following steps of: constructing a training data set by using an underexposure low-light image and an aligned low-light image in a plurality of scenes; constructing and training a low-light image denoising network in combination with a preset low-light loss function; the low-illumination loss function comprises loss for measuring the difference between two pairs of transformed low-light images, loss for measuring the difference between the denoised and down-sampled image and the target image, and structural similarity loss; constructing and training a decomposition network based on the de-noised image output by the LD-Net; the loss function of the decomposition network comprises reconstruction loss, reflectivity consistency loss, illumination smoothness loss and decomposition loss; and a de-noising Retinex decomposition model is generated according to the LD-Net and the decomposition network, and is used for carrying out image enhancement operation on the to-be-processed image based on the illumination correction factor, so that the model has less parameter quantity through a simple network design, and meanwhile, the model has better detail recovery effect and enhancement effect.

Description

Technical Field

[0001] The present application relates to the technical field of image enhancement, and more specifically, to a low-light image enhancement method. Background Art

[0002] As an important medium for information transmission, the quality of images has a crucial impact on subsequent downstream visual tasks. Therefore, image enhancement technology has gained wide attention and application in many application scenarios. In the field of computer vision research, low-light image enhancement technology is an important component. Its core goal is to enhance image quality by improving image brightness, adjusting contrast, improving color non-uniformity, suppressing noise and other operations, thereby providing clearer and more reliable image input for subsequent tasks.

[0003] With the deepening of research, low-light image enhancement technology has made significant progress. According to the different technical means adopted, the existing low-light image enhancement methods can be divided into two categories, namely traditional enhancement methods and deep learning-based enhancement methods. Traditional methods usually use manually designed prior knowledge to improve contrast and restore details. However, since low-light image enhancement is an ill-posed problem, it is difficult to adopt suitable priors under various lighting conditions, which limits the practical application of this method.

[0004] With the development of deep learning, many low-light image enhancement problems have been effectively solved. Instead of using manually designed prior knowledge, learning-based methods use deep neural networks to automatically learn features for low-light image enhancement. Thanks to its ability to model the mapping between low-light and high-quality images, deep learning methods usually achieve better results than traditional methods. Lore [Lore KG, Akintayo A.Sarkar S.LLNet: A deep autoencoder approach to natural low-light image enhancement [J]. Pattern Recognition, 2017, 61: 650-662] et al. first introduced deep learning into the low-light image enhancement task and proposed a stacked sparse denoising autoencoder (SSDA). This model effectively suppresses noise through sparse characteristics and achieves remarkable results. Subsequently, inspired by the multi-scale Retinex enhancement algorithm, Shen [Shen L, Yue Z, Feng F, et al. Msr-net: Low-light image enhancement using deep convolutional network [J]. arXiv preprint arXiv: 1711.02488, 2017] et al. first revealed the relationship between multi-scale Retinex and feed-forward convolutional neural network, and proposed MSR-Net. The network model includes three modules: multi-scale logarithmic transformation, convolution difference and color restoration. It achieved excellent performance by training with high-quality images and corresponding synthetic low-light images. Zhang [Zhang Y, Zhang J, Guo X. Kindling the darkness: A practical low-light image enhancer [C] / / Proceedings of the 27th ACM international conference on multimedia. 2019: 1632-1640] et al. proposed a KinD network, which includes an image decomposition module, a reflection restoration module, and an illumination adjustment module. The KinD network uses the illumination component to assist in reconstructing the reflection component, successfully alleviating the color distortion problem and improving the detail modeling capability.

[0005] Most of the existing low-light image enhancement methods rely on paired data of low-light images and normal-light images for supervised learning to capture the mapping relationship between brightness. However, since it is time-consuming and laborious to obtain paired datasets, how to reduce the dependence on paired datasets has become a key challenge in the field of low-light image enhancement. Summary of the invention

[0006] In view of this, the present application provides a low-light image enhancement method to solve the above-mentioned problem.

[0007] The technical solutions provided by this application are as follows:

[0008] A low-light image enhancement method, comprising:

[0009] Obtain N (N≥1) low-light image pairs to construct a training data set; each of the low-light image pairs consists of an underexposed low-light image and an aligned low-light image of the same scene;

[0010] Based on the training data set and the preset low-light loss function, a low-light image denoising network LD-Net is constructed and trained, and the LD-Net is used to output a denoised image; the low-light loss function includes a loss for measuring the difference between two pairs of transformed low-light images, a loss for measuring the difference between the denoised and downsampled image and the target image, and a structural similarity loss;

[0011] Based on the denoised image output by LD-Net, a decomposition network is constructed and trained; the loss function of the decomposition network includes reconstruction loss, reflectivity consistency loss, lighting smoothness loss and decomposition loss;

[0012] A denoising Retinex decomposition model is generated according to the trained LD-Net and the decomposition network, and is used to perform an image enhancement operation on the image to be processed based on the illumination correction factor.

[0013] In one possible implementation manner, after acquiring the low-light image pair, the method further includes:

[0014] The low-light image pair is preprocessed, and each low-light image in the low-light image pair is cropped to a preset size.

[0015] In one possible implementation, the constructed low-light denoising network LD-Net is a convolutional neural network CNN, including a convolutional layer Conv and an activation function layer added after the convolutional layer, and the activation function layer uses a linear correction unit LeakyReLU as an activation function;

[0016] The convolution layer Conv includes a 3x3 convolution kernel, which is used to extract local features of the low-light image input to LD-Net.

[0017] One possible implementation method is to train a low-light image denoising network LD-Net based on the training dataset, including:

[0018] Perform a downsampling operation G=(G1, G2) on the image I1 of size W×H×C in the training data set to obtain two images I1′=G1(I1) and I2′=G2(I1) of size W / 2×H / 2×C;

[0019] The low-light image denoising network LD-Net is trained based on I1′, I2′ and I1.

[0020] In one possible implementation, the denoising loss function preset by the low-light denoising network LD-Net is:

[0021] L LD =L Sym +L Con +L Ssim

[0022] Among them, L sym is a symmetric loss function, L Con is the loss function used to measure the difference between the denoised and downsampled image and the target image, L Ssim is the structural similarity loss function;

[0023] The symmetric loss function L Sym Used to measure the difference between two pairs of transformed low-light images, expressed as:

[0024]

[0025] In the formula, The image I1′ and the noise corresponding to the image I1′ are (Denoising) operation, The image I2′ is denoised by the noise corresponding to the image I2′. represents the L2 norm;

[0026] The loss function L Con Used to measure the two denoised and downsampled images i1′, i2′ and the corresponding expected or target image and The difference between is expressed as:

[0027]

[0028] The structural similarity loss function L Ssim It is expressed as:

[0029] L Ssim =1-SSIM(i1,i2)

[0030] Where SSIM(,) is a function that measures the structural similarity of two low-light images, i1 and i2 are the images corresponding to a low-light image pair after denoising.

[0031] In one possible implementation, the decomposition network includes an illumination component decomposition network IL-Net and a reflectance component decomposition network RE-Net, both of which are composed of six 3×3 convolutional layers with Relu as an activation function and one Sigmoid function.

[0032] In one possible implementation, the loss function of the decomposition network is set to:

[0033] L LR =L Rec +L Ref +L IL +L Dec

[0034] Among them, L Rec is the reconstruction loss, L Ref is the reflectivity consistency loss, L IL is the lighting smoothness loss, L Dec To decompose the loss.

[0035] The reconstruction loss function is used to ensure that the decomposed reflectance component R and illumination component L can reconstruct the original input denoised image i, which is expressed as:

[0036]

[0037] Where ⊙ is the element-by-element multiplication operation;

[0038] Reflectivity consistency loss function The consistency of the reflectivity map under different lighting conditions is expressed as:

[0039]

[0040] The lighting smoothness loss can be expressed as:

[0041]

[0042] In the formula, and Represent the first-order derivatives in the horizontal and vertical directions respectively, ||||1 is the L1 norm;

[0043] The decomposition loss function is expressed as:

[0044]

[0045] Where stopgrad(L) is the gradient at which the illumination component is stopped from being updated, L0 is the initial illumination estimate, and L0 = maxI C (x), where C∈{R,G,B}.

[0046] One possible implementation method is to generate a denoising Retinex decomposition model and perform image enhancement operations on the image to be processed based on the illumination correction factor, including:

[0047] The image to be processed is input into the denoising Retinex decomposition model, and image enhancement is performed based on the low-light image denoising network and the decomposition network, which is expressed as:

[0048] I Fin =g(L)⊙R=L λ ⊙R

[0049] Among them, g(L) is the Gamma correction, ⊙ is the element-by-element multiplication operation, λ is the illumination correction factor, I Fin Represents the final enhanced image.

[0050] Compared with the prior art, the beneficial effects of this application are:

[0051] In the image enhancement process, the present application first applies a denoising network (LD-Net) to denoise paired low-light images, which effectively prevents detail loss and color deviation, thereby retaining more details and more accurate color representation in the enhanced image. In the image decomposition stage after denoising, by maintaining the consistency of the reflectance component, the recovery of details is further ensured, which helps to maintain the visual consistency of the image under different lighting conditions and enhance the realism of the image. By applying the illumination correction factor to adjust the illumination component, the present application can achieve more accurate illumination correction, thereby providing better visual effects when enhancing images. The introduction of the illumination correction factor makes the present application more flexible and adaptable when processing images under different lighting conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 A flowchart of a low-light image enhancement method provided in Example 1 of the present application.

[0053] Figure 2 A schematic diagram of the collection of multiple scenarios for constructing a training data set and the distribution of their sample numbers provided in Example 1 of the present application.

[0054] Figure 3 A schematic diagram of constructing and training a low-light image denoising network provided in Example 1 of the present application.

[0055] Figure 4 A schematic diagram of the structure of the decomposed network provided in Example 1 of the present application.

[0056] Figure 5 A flowchart of image enhancement processing based on a denoising Retinex decomposition model is provided in Example 1 of the present application. DETAILED DESCRIPTION

[0057] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present application.

[0058] Embodiment 1

[0059] Embodiment 1 of the present application provides a low-light image enhancement method based on a denoising Retinex decomposition network. A number of low-light images are collected as training data, and a denoising Retinex decomposition model with denoising and detail restoration capabilities is obtained through training to achieve low-light image enhancement.

[0060] The low-light image enhancement method provided by this application is described in detail below in conjunction with specific implementation steps. Figure 1 , is a flow chart of a low-light image enhancement method provided in Example 1 of the present application. The above method specifically includes the following steps:

[0061] Step 101: extract N (N≥1) low-light image pairs to construct a training data set.

[0062] Each low-light image pair consists of an underexposed low-light image and a well-aligned low-light image of the same scene. In some embodiments, underexposed and well-aligned low-light images can be directly extracted from a public data set, such as using the SID data set, which provides registered RAW image pairs of the same scene. In each pair of images, one image has a short exposure time and insufficient exposure time, and the other image has a long exposure time and good exposure. By comparing the exposure time information and brightness histogram of each image in each image pair when it is taken, the image with a shorter exposure time and lower brightness is selected as the underexposed low-light image.

[0063] In addition, considering that many machine learning and deep learning models, especially convolutional neural networks (CNNs), require input data to have a uniform size to ensure that all images have the same dimension, this embodiment uniformly crops the images in each extracted low-light image pair to a preset size to obtain a training data set after preprocessing.

[0064] In this embodiment, underexposed and well-aligned low-light image pairs are selected from the public dataset to construct the training dataset. Finally, 324 scenes are collected, each scene includes 2-6 images, and a total of 1000 images. Figure 2As shown in , the number of images contained in each scene is shown. In the figure, the X-axis indicates that there are 324 different data points in the data set. Each data point corresponds to a scene. The Y-axis indicates the number of samples. The number of images in each scene collected in this embodiment varies between 2 and 6. Compared with other state-of-the-art unsupervised learning methods, after testing on multiple data sets, it has better detail recovery and enhancement effects.

[0065] During training, two low-light images are randomly selected in each scene to form a low-light image pair, and each low-light image is cropped to the preset 128×128, and the input is set to a 128×128 low-light image pair.

[0066] Step 102: Construct a low-light image denoising network LD-Net.

[0067] like Figure 3 As shown in , the low-light image denoising network constructed in this application adopts a CNN structure, including a convolution layer Conv and an activation function layer. The convolution layer Conv is used to extract local features of the low-light image input to the low-light denoising network. Each convolution layer Conv uses a 3x3 convolution kernel, that is, each convolution kernel covers 3 pixels wide and 3 pixels high, and can detect edges, textures and other local features in the image. Each convolution layer has 48 output channels (feature maps) that can extract 48 different features of the low-light image input to the low-light denoising network. And, as Figure 3 As shown in , in consecutive convolutional layers, the number of output channels of the previous layer matches the number of input channels of the next layer. That is, if the first convolutional layer has 48 output channels, then the second convolutional layer also contains 48 input channels to ensure the correct flow of data.

[0068] After each convolutional layer, the low-light image denoising network LD-Net uses the LeakyRectified Linear Unit (LeakyReLU) as the activation function, so that the low-light image denoising network can learn and simulate complex function mapping relationships.

[0069] Selecting LeakyReLU as the activation function can solve the problem of the traditional ReLU activation function having a zero gradient when the input is negative, which is the so-called "neuron death" problem. In LeakyReLU, even if the input is negative, there will be a small positive slope (usually 0.01), so that the gradient is not zero, allowing the low-light image denoising network to learn even with negative input values. At the same time, LeakyReLU also retains negative feature information, which helps improve the training efficiency and generalization ability of the model.

[0070] After the convolutional layers and activation functions, the low-light image denoising network is able to generate feature maps, which represent the feature representations of the low-light image input to the network at different levels. Figure 3 As shown in , the number of feature maps corresponds to the number of convolution kernels in the convolution layer, and a total of 9 feature maps can be generated.

[0071] Step 103: Train a low-light image denoising network LD-Net based on the above training data set.

[0072] Specifically, the image I1 of size W×H×C in the training dataset is downsampled by G=(G1, G2) to obtain two images I1′=G1(I1) and I2′=G2(I1) of size W / 2×H / 2×C. This operation divides I1 into non-overlapping 2×2 pixel blocks, I1′ takes the average of diagonal pixels, and I2′ takes the average of anti-diagonal pixels. Subsequently, I1′, I2′ and I1 are used to train the low-light image denoising network.

[0073] like Figure 3 As shown in , image I1, I1′ and I2′ obtained after downsampling are used as inputs of the low-light image denoising network. After processing images I1, I1′ and I2′ through multiple convolutional layers and activation function layers of the low-light image denoising network, 9 feature maps corresponding to image I1 can be obtained.

[0074] During the training process of the low-light image denoising network, a reasonable low-light loss function must be set to ensure the accuracy of denoising. The low-light loss function used by the low-light image denoising network in this embodiment can be expressed as:

[0075] L LD =L Sym +L Con +L Ssim

[0076] Among them, L Sym is a symmetric loss function, L Con is the loss function used to measure the difference between the denoised and downsampled image and the target image, L Ssim is the structural similarity loss function.

[0077] Symmetric loss function L Sym It is used to measure the difference between two pairs of transformed low-light images and can be expressed as:

[0078]

[0079] In the formula, The image I1′ and the noise corresponding to the image I1′ are (Denoising) operation, For image I2′ and the noise corresponding to image I2′ The (denoising) operation is performed, that is, different noises are added to image I1′ and image I2′ respectively, and then the original image is estimated or restored by training a low-light denoising network. Represents the L2 norm.

[0080] By minimizing the symmetric loss function L Sym , the model can learn how to make Closer to I2′, while making Closer to I1′, or making the difference between the two pairs consistent.

[0081] The above loss function L Con It can be expressed as:

[0082]

[0083] Image / noise obtained through low-light image denoising network The denoising operation is performed to obtain the denoised image i, and then the denoised image i is downsampled by G = (G1, G2) to obtain l1 and l2. The loss function L Con Used to measure the two denoised and downsampled images i1′, i2′ and the corresponding expected or target image and The difference between.

[0084] By minimizing L Con , the model can learn how to effectively remove noise and maintain the key features of the image, while reducing the amount of data through downsampling operations, which helps to reduce computational and storage requirements while maintaining image quality.

[0085] Structural similarity loss function L Ssim , which can be expressed as:

[0086] L Ssim =1-SSIM(i1,i2)

[0087] Where SSIM(,) is a function that measures the structural similarity of two low-light images, i1 and i2 are the images corresponding to a low-light image pair after denoising.

[0088] Step 104: training a decomposition network based on the denoised low-light image output by the low-light image denoising network.

[0089] The decomposition network constructed in this embodiment adopts a CNN structure, which includes an illumination component decomposition network IL-Net and a reflectivity component decomposition network RE-Net. Figure 4As shown in , both IL-Net and RE-Net contain seven convolutional layers. The first six layers are 3×3 convolutional layers with Relu as the activation function, and the last layer uses the Sigmoid function to constrain the output to the range of [0, 1].

[0090] For the illumination component, it is assumed that the three RGB channels have the same value. Therefore, the number of IL-Net output channels is set to 1 and the number of RE-Net channels is set to 3. The low-light denoised images i1 and i2 are used as the input of the decomposition network respectively.

[0091] The loss function of the decomposition network is set as:

[0092] L LR =L Rec +L Ref +L IL +L Dec

[0093] Among them, L Rec is the reconstruction loss, L Ref is the reflectivity consistency loss, L IL is the lighting smoothness loss, L Dec To decompose the loss.

[0094] The reconstruction loss function can be expressed as:

[0095] It is used to ensure that the decomposed reflectance component R and illumination component L can reconstruct the original input denoised image i. Where ⊙ is an element-by-element multiplication operation.

[0096] The reflectivity consistency loss function can be expressed as:

[0097] Based on Retinex theory, paired low-light images should have the same reflectance because they share the same object, and the reflectance consistency loss function restricts the consistency of reflectance maps under different lighting conditions.

[0098] The lighting smoothness loss can be expressed as:

[0099]

[0100] In the formula, and denote the first-order derivatives in the horizontal and vertical directions, respectively, and ||||1 is the L1 norm. In natural scenes, illumination is often uniformly distributed, while different materials and surfaces reflect light differently, which may lead to variations in brightness. Therefore, the illumination is assumed to be piecewise smooth in order to better capture these variations in natural scenes.

[0101] The decomposition loss function is expressed as:

[0102]

[0103] Based on the above decomposition loss function to guide the decomposition, the illumination is first estimated and the reflectance is calculated by pixel-by-pixel division of image i and its illumination component L.

[0104] Where stopgrad(L) is to stop updating the gradient of the illumination component to stabilize the training. In Retinex theory, the maximum value L0 of the R, G, and B channel intensities is usually used as the initial illumination estimate, L0 = maxI C (x), where C∈{R,G,B}.

[0105] Step 105: Generate a denoising Retinex decomposition model based on the trained LD-Net and decomposition network, so as to perform image enhancement operation on the image to be processed.

[0106] Specifically, the trained denoising Retinex decomposition model has already had quite good denoising and detail restoration capabilities. Figure 5 As shown in , the low-light image denoising network and the decomposition network are applied to calculate the final enhanced image, which is expressed as:

[0107] I Fin =g(L)⊙R=L λ ⊙R

[0108] Where g(L) is the Gamma correction, ⊙ is the element-by-element multiplication operation, λ is the illumination correction factor, and I Fin Represents the final enhanced image.

[0109] Compared with the prior art, the technical solution provided by the embodiment of the present application has the following beneficial effects:

[0110] This application uses a convolutional neural network to build a network model, designs a low-light image denoising network and a decomposition network, and the simpler network design makes the model have fewer parameters. In the image enhancement process, the denoising network (LD-Net) is first applied to denoise the paired low-light images, which effectively prevents detail loss and color deviation, thereby retaining more details and more accurate color representation in the enhanced image. In the image decomposition stage after denoising, by maintaining the consistency of the reflectivity component, the restoration of details is further ensured, which helps to maintain the visual consistency of the image under different lighting conditions and enhance the realism of the image. By applying the illumination correction factor to adjust the illumination component, the present application can achieve more accurate illumination correction, thereby providing better visual effects when enhancing the image. The introduction of the illumination correction factor makes the present application more flexible and adaptable when processing images under different lighting conditions.

[0111] Although the embodiments of the present application have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present application, and that the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A low-light image enhancement method, characterized in that: include: Obtain N (N≥1) low-light image pairs to construct a training dataset; Each of the low-light image pairs consists of an underexposed low-light image and an aligned low-light image of the same scene; Based on the training data set and the preset low-light loss function, a low-light image denoising network LD-Net is constructed and trained, and the LD-Net is used to output a denoised image; the low-light loss function includes a loss for measuring the difference between two pairs of transformed low-light images, a loss for measuring the difference between the denoised and downsampled image and the target image, and a structural similarity loss; Based on the denoised image output by LD-Net, a decomposition network is constructed and trained; the loss function of the decomposition network includes reconstruction loss, reflectivity consistency loss, lighting smoothness loss and decomposition loss; A denoising Retinex decomposition model is generated according to the trained LD-Net and the decomposition network, and is used to perform an image enhancement operation on the image to be processed based on the illumination correction factor.

2. The low-light image enhancement method according to claim 1, characterized in that: After acquiring the low-light image pair, the method further includes: The low-light image pair is preprocessed, and each low-light image in the low-light image pair is cropped to a preset size.

3. The low-light image enhancement method according to claim 1, characterized in that: The constructed low-light denoising network LD-Net is a convolutional neural network CNN, including a convolutional layer Conv and an activation function layer after the convolutional layer is added, and the activation function layer uses a linear correction unit LeakyReLU as an activation function; The convolution layer Conv includes a 3x3 convolution kernel, which is used to extract local features of the low-light image input to LD-Net.

4. The low-light image enhancement method according to claim 1, characterized in that: The low-light image denoising network LD-Net is trained based on the training dataset, including: Perform a downsampling operation G=(G1, G2) on the image I1 of size W×H×C in the training data set to obtain two images I1′=G1(I1) and I2′=G2(I1) of size W / 2×H / 2×C; The low-light image denoising network LD-Net is trained based on I1′, I2′ and I1.

5. The low-light image enhancement method according to claim 1, characterized in that: The denoising loss function preset by the low-light denoising network LD-Net is: L LD =L Sym +L Con +L Ssim Among them, L sym is a symmetric loss function, L Con is the loss function used to measure the difference between the denoised and downsampled image and the target image, L Ssim is the structural similarity loss function; The symmetric loss function L Sym Used to measure the difference between two pairs of transformed low-light images, expressed as: In the formula, The image I1′ and the noise corresponding to the image I1′ are (Denoising) operation, The image I2′ is denoised by the noise corresponding to the image I2′. represents the L2 norm; The loss function L Con Used to measure the two denoised and downsampled images i1′, i2′ and the corresponding expected or target image and The difference between is expressed as: The structural similarity loss function L Ssim It is expressed as: L Ssim =1-SSIM(i1,i2) Where SSIM(,) is a function that measures the structural similarity of two low-light images, and i1 and i2 are the images corresponding to a low-light image pair after denoising.

6. The low-light image enhancement method according to claim 1, characterized in that: The decomposition network includes an illumination component decomposition network IL-Net and a reflectance component decomposition network RE-Net, both of which are composed of six 3×3 convolutional layers with Relu as the activation function and one Sigmoid function.

7. The low-light image enhancement method according to claim 1, characterized in that: The loss function of the decomposition network is set as: L LR =L Rec +L Ref +L IL +L Dec Among them, L Rec is the reconstruction loss, L Ref is the reflectivity consistency loss, L IL is the lighting smoothness loss, L 0ec To decompose the loss. The reconstruction loss function is used to ensure that the decomposed reflectance component R and illumination component L can reconstruct the original input denoised image i, which is expressed as: Where ⊙ is the element-by-element multiplication operation; Reflectivity consistency loss function The consistency of the reflectivity map under different lighting conditions is expressed as: The lighting smoothness loss can be expressed as: In the formula, and Represent the first-order derivatives in the horizontal and vertical directions respectively, || ||1 is the L1 norm; The decomposition loss function is expressed as: Where stopgrad(L) is the gradient at which the illumination component is stopped from being updated, L0 is the initial illumination estimate, and L0 = maxI c (x), where C∈{R,G,B}.

8. The low-light image enhancement method according to claim 1, characterized in that: Generate a denoising Retinex decomposition model and perform image enhancement operations on the image to be processed based on the illumination correction factor, including: The image to be processed is input into the denoising Retinex decomposition model, and image enhancement is performed based on the low-light image denoising network and the decomposition network, which is expressed as: I Fin =g(L)⊙R=L λ ⊙R Among them, g(L) is the Gamma correction, ⊙ is the element-by-element multiplication operation, λ is the illumination correction factor, I Fin Represents the final enhanced image.

Images