A low-light image enhancement method based on parameter estimation

Abstract

The application discloses a kind of faint light image enhancement methods based on parameter estimation, belong to computer vision, machine learning and other fields.Utilize depth separable convolution and the parameter of optimization network of reverse residual structure, the ability of network to the extraction of channel and spatial dimension feature is guaranteed, while the calculation amount and parameter amount of convolution operation are reduced.Then, the scaling point attention mechanism that can capture global information of feature is used to estimate color correction matrix and gamma correction parameter, realizes the overall optimization to final imaging quality.In the network training process, the L2 loss of pixel-level difference and the perception loss of multi-weight cognitive-level difference are used.The experiment of the application on LoL dataset shows that when the expansion rate of reverse residual structure is 1.5, the optimal performance can be achieved on LoL dataset.

Description

Technical Field

[0001] This invention belongs to the fields of computer vision and machine learning, and specifically relates to a low-light image enhancement method based on parameter estimation. Background Technology

[0002] Low-light image enhancement is a common image processing task aimed at improving image visibility while maintaining its natural appearance. It effectively addresses issues such as insufficient lighting and uneven illumination, thereby improving image quality and making it easier to extract relevant or visually consistent features. Low-light image enhancement is often used as a preprocessing technique for computer vision tasks such as object recognition and detection, helping these tasks achieve better results and accuracy. Low-light image enhancement tasks can be categorized into traditional enhancement methods and deep learning-based methods, depending on the processing approach used.

[0003] Traditional image enhancement methods are mainly divided into non-physical model methods and physically-based methods. Non-physical model methods are an important image processing technique that enhances images by directly adjusting image pixels to improve visibility and quality. Physically-based methods primarily model the degradation process of low-light images, continuously adjusting imaging parameters to ultimately reconstruct the undegraded image. However, due to limitations in feature definition and processing, traditional enhancement methods are difficult to apply in real-world scenarios.

[0004] With the rapid development of deep learning, more and more low-light image enhancement methods have adopted deep learning-related methods. Deep learning low-light image enhancement methods can be divided into four types: supervised, unsupervised, semi-supervised and self-supervised. Among them, supervised and unsupervised methods are currently the most widely used methods, which can be summarized as follows: (1) Supervised methods use paired datasets to train enhancement networks. For example, Zhang et al. proposed the KIND enhancement method, which decomposes the image into incident and reflected components, further optimizes and corrects these two components, and finally linearly combines them into the final enhanced image; Lv et al. proposed MBLLEN, which implicitly distributes the noise reduction, brightness enhancement and color restoration in the image enhancement process to each branch through multi-channel brightness enhancement. (2) Unsupervised methods set the enhancement target and guide the presentation effect of the enhanced image through different loss functions. For example, Guo et al. designed the Zero-DC method, which regards the light enhancement task as a customized mapping curve and designed a reference-free loss function and parameter estimation network DCE-net to enhance the image, which has the characteristics of high efficiency. Jiang et al. adopted an adversarial generative approach, combining U-net with attention as the generator and designing a discriminator. They used the set adversarial loss and the designed natural feature preservation loss to achieve natural effect enhancement of low-light images. (3) Semi-supervised methods are trained using labeled and unlabeled data. For example, Yang et al. first trained the designed enhancement network on paired data with multi-level content loss to recover image details. Subsequently, they used a domain adaptation mechanism to train the network on unpaired real data with adversarial training and a set loss for gamma correction to improve the performance of the network model. (4) Self-supervised methods are achieved by directly mining information from the data. For example, Zhang et al., based on the Retinex theory of maximum entropy, combined the enhancement network of the self-supervised method, constraining the channel value with the highest reflectivity to be equal to the maximum channel value of the low-light image, and the entropy value of the channel with the highest reflectivity is also the largest, and trained the network accordingly.

[0005] Low-light image enhancement often requires addressing both brightness and noise reduction. However, unsupervised enhancement methods struggle to achieve both goals simultaneously. Therefore, this invention follows a supervised training approach and proposes a low-light image enhancement method based on estimating transformed image parameters.

[0006] This invention reveals several problems with current enhancement methods, requiring further optimization. Three significant issues stand out: severe noise, color distortion, and excessively large model parameters. Severe noise often arises from the low signal-to-noise ratio of low-light images, where enhancement methods amplify the existing noise while increasing brightness. The color balance of low-light images may be biased towards a particular color channel. To improve overall brightness, enhancement often requires boosting the brightness of all channels, potentially leading to over-enhancement of certain color channels. Without proper color correction or with ineffective correction, color distortion will occur in the enhanced image. Low-light image enhancement models must simultaneously consider both enhancing visual quality and preserving original image information, necessitating complex image processing such as brightness transformation, noise reduction, and color modification. This necessitates designing model structures with numerous layers and parameters to estimate more information, resulting in an excessively large number of model parameters.

[0007] To address the above problems, this invention proposes a parameter estimation-based enhancement network. The following sections will discuss the structure and function of the modules used in the network design process, and experiments will demonstrate the effectiveness of the proposed enhancement algorithm in low-light image enhancement. Summary of the Invention

[0008] The purpose of this invention is to propose a low-light image enhancement method based on parameter estimation. This method is implemented through a low-light image enhancement network based on parameter estimation. The network consists of a pixel-level component estimation branch and an overall optimization branch, achieving pixel-level brightness adjustment, noise reduction, and overall color and brightness optimization of the image. Specifically, the pixel-level component estimation branch calculates the components for adjusting brightness and mitigating noise, while the overall optimization branch calculates the color correction matrix and gamma correction parameters. To facilitate verification of the proposed method, it was trained using the same dataset as other low-light image enhancement methods, namely the Low-Light Pairing Dataset (LOL dataset). After training, testing was conducted on the corresponding test set within this dataset. Comparison revealed that the proposed method effectively enhances low-light images. Furthermore, comparative experiments with other methods demonstrate the effectiveness of the proposed method.

[0009] To achieve the above objectives, this invention proposes a low-light image enhancement method based on parameter estimation.

[0010] This method employs a low-light image enhancement network to enhance the input low-light image; the low-light image enhancement network includes: a pixel-level component estimation branch and an overall optimization branch;

[0011] The pixel-level component estimation branch includes a first branch and a second branch with identical structures. The first or second branch, from input to output, sequentially includes: a first convolutional module, N inverse residual structures, and a second convolutional module. The input of the additional second convolutional module is a residual concatenation of the output of the first convolutional module and the outputs of the N inverse residual structures. The input of the first branch is the input of the low-light image enhancement network, and the input of the second branch is the result of a dot product between the output of the first branch and the input of the low-light image enhancement network, followed by concatenation with the input of the low-light image enhancement network. The output of the pixel-level component estimation branch is the sum of the output of the first branch and the output of the second branch.

[0012] The overall optimization branch includes: a third convolutional module, a fourth convolutional module, a first fully connected layer, a second fully connected layer, a scaling attention module, a third fully connected layer, layer normalization, a fourth fully connected layer, an activation function, a fifth fully connected layer, a first convolutional layer, and a second convolutional layer. The input to the overall optimization branch is the concatenation of the input to the low-light image enhancement network and the output of the pixel-level component estimation branch. The input to the overall optimization branch passes through the third and fourth convolutional modules sequentially and is then output to the first and second fully connected layers, respectively. The input to the scaling attention module includes: the output of the first fully connected layer, the output of the second fully connected layer, and a query parameter. This query parameter is used to limit the output value of the overall optimization branch. The output of the scaling attention module sequentially passes through the third fully connected layer, layer normalization, the fourth fully connected layer, the activation function, and the fifth fully connected layer. The output of the third fully connected layer is residually concatenated with the output of the fifth fully connected layer and then added to a 3x3 identity matrix and a value of size 1. The result of the addition is split into a 3x3 matrix and a value of size 1. The 3x3 matrix is ​​input to the first fully connected layer, and the value of size 1 is input to the second fully connected layer. The output of the overall optimization branch is the result of matrix multiplication of the output of the pixel-level component estimation branch with the output of the first fully connected layer and then gamma transformation of the output of the second fully connected layer, which is the output of the low-light image enhancement network.

[0013] The loss of the low-light image enhancement network as follows:

[0014]

[0015]

[0016]

[0017] in, The loss represents the pixel difference between the predicted image obtained by the low-light enhancement network and the corresponding high-brightness image in the LOL dataset. This represents the corresponding high-brightness image in the LOL dataset. The image represents the predicted image obtained by the low-light enhancement network. , , These represent the height, width, and number of channels of the real image, respectively, with superscript indicating the height, width, and number of channels. This represents the coordinates of the corresponding image feature in RGB space. , , Representing VGG-16 The height, width, and number of channels of the layer output features. This refers to the sequence number of the VGG-16 network layer. Representing the VGG-16 network The output of the layer, This represents the perceptual loss obtained using VGG-16.

[0018] Furthermore, the first and second convolutional modules in the pixel-level component estimation branch have the same structure, both including: a convolutional layer, a batch normalization layer, and an activation function in sequence; the inverse residual structure includes: a first pointwise convolutional module, a channelwise convolutional module, and a second pointwise convolutional module. The input of the first pointwise convolutional module is the input of the inverse residual structure, and the output of the inverse residual structure is the residual connection value between the output of the second pointwise convolutional module and the input of the inverse residual structure.

[0019] Furthermore, the number N of inverted residual structures in the first or second branch of the pixel-level component estimation branch is 6.

[0020] This invention proposes using a convolutional structure that ensures consistent spatial feature resolution to learn the adjustment components of pixel-level images, preserving the detailed features of the original images and utilizing the contextual information between pixel positions in the images. Simultaneously, depthwise separable convolutions and inverse residual structures are used to optimize the network parameters, ensuring the network's ability to extract channel and spatial dimension features while reducing the computational cost and parameter count of convolution operations. Then, a scaling-point attention mechanism that captures global feature information is used to estimate the color correction matrix and gamma correction parameters, achieving overall optimization of the final image quality. During network training, this invention uses L2 loss for pixel-level differences and multi-weighted perceptual loss for cognitive-level differences. Experiments on the LoL dataset show that an inverse residual structure with a dilation rate of 1.5 achieves optimal performance on the LoL dataset. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the low-light enhancement network structure of the present invention.

[0022] Figure 2This is a comparison chart of the effects of this invention on the LoL dataset with other low-light image enhancement methods both domestically and internationally.

[0023] Figure 3 This is a comparison chart of the effects of this invention on the Exdark dataset with other low-light image enhancement methods both domestically and internationally. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings, so that those skilled in the art can better understand this invention. It should be particularly noted that the described embodiments are only some embodiments of this invention, not all embodiments, and are not intended to limit the scope of the claimed invention. All other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this invention.

[0025] The following provides a detailed explanation of each of the details involved in the technical solution of this invention.

[0026] This invention proposes a low-light image enhancement method based on parameter estimation, which includes the following steps:

[0027] (1) Design a low-light image enhancement network

[0028] The overall structure of the low-light image enhancement network is as follows: Figure 1 As shown. Inspired by the Zero-DCE method, the low-light image enhancement network proposed in this invention aims to estimate a parameter component with the same size as the input low-light image that can adaptively adjust brightness, denoted as the low-light adjustment component, to achieve brightness enhancement of the low-light image. Simultaneously, addressing the issue that Zero-DCE does not consider the influence of low-light image noise, this invention considers that most noise in the input low-light image is presented in a superimposed manner. Therefore, an offset, denoted as the noise mitigation component, is designed based on the low-light adjustment component to minimize the interference of low-light image noise, while also learning the superimposed brightness to a certain extent. Multiplying and adding the input low-light image with the low-light adjustment component and the noise mitigation component yields a preliminary normal illumination image. The enhancement formula is as follows:

[0029]

[0030] in, This is a preliminary image under normal lighting conditions. It is a low-light image. It is a low-light adjustment component with the same size as the input low-light image. It is a noise reduction component with the same size as the input low-light image.

[0031] To further optimize the preliminary normal illumination image obtained from pixel variations, the low-light image enhancement network of this invention also improves the quality of the preliminary normal illumination image from the perspective of overall imaging by estimating the parameters of the color correction matrix and gamma correction. The optimization formulas using the color correction matrix and gamma correction parameters are as follows:

[0032]

[0033] in, This represents a color transformation operation used to improve the colors of a pre-existing, normally lit image. Gamma correction is used to optimize and enhance dark areas in an image. Since gamma correction adjusts brightness globally, it can also smooth out differences in brightness across the image. Represents a preliminary image under normal lighting conditions. Optimized, high-quality image.

[0034] In constructing the pixel-level component estimation branches for estimating the low-light adjustment component and the noise mitigation component, this invention introduces a high-performance convolutional structure to further optimize the number of parameters in the pixel-level component estimation branches, thereby reducing the size of the network model and ensuring its efficient operation. For estimating the gamma correction parameters of the color correction matrix and gamma correction, this invention borrows from scaling point attention to design an overall optimized branch structure.

[0035] For an input low-light image Micro-light adjustment components estimated by pixel-level component estimation branch and noise mitigation components Generate a preliminary image under normal lighting conditions. Then, the color transformation operation of the branch was optimized as a whole. and gamma correction operation Generate the final high-quality image after low-light enhancement.

[0036] (2) Implementation of the network structure for the pixel-level component estimation branch

[0037] The pixel-level component estimation branch is used to estimate the low-light adjustment component and the noise mitigation component. To preserve the detailed features of the input low-light image, unlike the downsampling and upsampling structures of previous encoders and decoders, the proposed pixel-level component estimation branch structure is based on the pixel-level branch in UIEC^2-Net, ensuring consistent spatial feature resolution. Meanwhile, low-light image enhancement tasks often require deployment on mobile devices, making it necessary to reduce the number of parameters to improve model scalability. To further reduce the number of parameters in the pixel-level component estimation branch while maintaining certain performance, this invention further optimizes the convolutional structure in the network through depthwise separable convolutions and inverse residual structures to pursue better performance.

[0038] Depthwise separable convolution separates features according to channel and spatial dimensions through channel-wise and pointwise convolutions. Features in different channel and spatial dimensions are extracted separately through convolution, and then concatenated to form the overall feature set. This ensures the integrity of the transformed feature space while reducing the number of parameters in the pixel-level component estimation branch. To further enhance the feature extraction capability of the pixel-level component estimation branch, an inverse residual structure is employed. This involves increasing the dimensionality through pointwise convolutions, extracting features within each channel through channel-wise convolutions, and finally reducing the dimensionality through pointwise convolutions again. This places feature extraction in a higher-dimensional space. Simultaneously, the lightweight nature of channel-wise and pointwise convolutions ensures a minimal number of parameters in the pixel-level component estimation branch model. Furthermore, the inverse residual structure enriches the input feature information, enhancing the feature extraction capability of the pixel-level component estimation branch network.

[0039] Because the pixel-level component estimation branch needs to estimate the low-light adjustment component and the noise mitigation component. Figure 1 In constructing the pixel-level component estimation branch, the number of channels in the feature space of the input low-light image is first expanded by convolution, and then fed into a feature extraction structure built by inverse residual blocks. To make training more stable, a residual connection edge is added, and then the features are compressed by convolution to obtain the low-light adjustment component. Next, the input low-light image processed by the low-light adjustment component and the original input low-light image are concatenated (the feature tensor is concatenated by channel dimension), and the noise reduction component is obtained through the feature extraction structure built by inverse residual blocks. Finally, the corresponding positions of the input low-light image and the low-light adjustment component are multiplied to obtain the multiplied image. Then, the multiplied image is iteratively added to the noise reduction component to obtain the preliminary enhanced image.

[0040] (3) Implementation of overall optimized branch network structure

[0041] The overall optimization branch of the network structure is used to estimate the color correction matrix and gamma correction parameters. For low-light image enhancement technology, the enhanced image often suffers from color deviation compared to the input low-light image. Furthermore, in real-world images, the presented effect often conforms to the human visual system, with inconsistent enhancement of bright and dark details, and the enhanced image often appears insufficient in color processing. Therefore, to further correct the color of the enhanced image and adjust the brightness of bright and dark details, this invention designs an overall optimization branch for processing the initial normal lighting image. By estimating the color correction matrix and gamma correction parameters, the final imaging effect is improved. Specifically, the color correction matrix is ​​determined by decomposing the relationship between the three RGB channels and then adjusting the channel values, a more reasonable approach. Gamma correction is used to match the non-linear perception of dark areas by humans, further optimizing the overall brightness of the image. Therefore, this invention constructs the overall optimization branch based on the following formula.

[0042]

[0043] in, This represents the pixel values ​​of the optimized, normally lit image output by the overall optimization branch. This represents the overall color transformation matrix, where γ represents the gamma correction parameter. This represents the input pixel values ​​of a preliminary, normally lit image. This represents the minimum setpoint value for the corrected pixel value. In the design of estimating the color correction matrix and gamma correction parameters, inspired by the Illumination Adaptive Transformer (IAT), parameter estimation is performed using scaling point attention, and parameters are calculated by capturing global features of the input features of the overall optimization branch. For example... Figure 1 As shown, this invention transforms the input features of the overall optimization branch into a series of key-value pairs in the scaling point attention mechanism through convolution and fully connected operations. The gamma correction parameters and color matrix parameters are used as the query. The correlation between the query value and a series of key values ​​in the scaling point attention mechanism is calculated. The final attention value of the query parameters in the scaling point attention mechanism is obtained by weighting the query with the weights of the value values ​​and the correlation between the query and the key values, using the weights of the value values.

[0044] Since the overall optimization branch only needs to estimate the gamma coefficients and color correction matrix for overall image adjustment, without considering the relative positions of pixels or local details, it is easier to generate features with global information by compressing the feature resolution. Therefore, for the features of the initial normal illumination image and the input low-light image Concat from the pixel-level component estimation branch, the image information is first encoded into a high-dimensional space through two convolutions, generating features with smaller resolution, more channels, and higher semantic levels. Then, the dimensionality of the features is compressed through two fully connected layers to generate the feature size required for scaling point attention. Next, the higher-level semantic features obtained from the convolution are fed into the scaling point attention module, where parameters are learned through the overall relationship of image features, and further optimized through fully connected layers to obtain the final gamma correction parameters and color correction matrix.

[0045] (4) Train the low-light image enhancement network model and verify its effectiveness through experiments.

[0046] The preceding steps completed the design of the enhancement network. Next, this invention requires training the low-light image enhancement network model. The goal of this invention's low-light image enhancement network is to make the final enhanced image as consistent as possible with the corresponding high-brightness image in the LOL dataset, while also being similar in feature space and image structure. Therefore, the overall loss for training the low-light image enhancement network is as follows:

[0047]

[0048]

[0049]

[0050] in, The loss represents the pixel difference between the predicted image obtained by the low-light enhancement network and the corresponding high-brightness image in the LOL dataset. This represents the perceptual loss obtained using VGG-16. This represents the total loss. This represents the corresponding high-brightness image in the LOL dataset. The image represents the predicted image obtained by the low-light enhancement network. , , These represent the height, width, and number of channels of the real image, respectively. Representing the VGG-16 network The output of the layer, This refers to the sequence number of the VGG-16 network layer. , , Representing VGG-16 The height, width, and number of channels of the layer output features.

[0051] Regarding the parameter design of the network structure, experiments revealed that the optimal inverse residual rate for the pixel-level component estimation branch of this invention is 1.5, and the final results are also based on an inverse residual rate of 1.5. The training dataset used was a pairwise LoL dataset. When training the low-light image enhancement network, the dataset was partitioned according to the standard LoL dataset partitioning method, the learning rate was set to 0.0004, the Adam optimizer was used, and the network was trained iteratively for 250 epochs.

[0052] Table 1. Comparison of objective metrics SSIM and PSNR between domestic and international methods on the LoL dataset.

[0053]

[0054] This invention will compare the proposed low-light image enhancement method with other existing low-light enhancement methods. Figure 2 Table 1 shows the test results on the LoL dataset. Regarding metrics, Table 1 shows that the augmentation method of this invention achieves a peak signal-to-noise ratio (PSNR) of 22.56 and a structural similarity (SSIM) of 0.807 on the LoL dataset. Among all methods, the augmentation method proposed in this invention performs best in SSIM and is only slightly inferior to the RCT method in terms of PSNR. Figure 2 The image enhancement effect comparison shown in the figure demonstrates that the enhancement method proposed in this invention not only most closely approximates the real highlight image in terms of color reproduction, but also significantly improves brightness and retains more details, with only a slight white cast. In contrast, although the MELLEN method performs well in color reproduction, it is ineffective in handling light and shadow; while the KIND method significantly improves brightness, its enhancement effect results in more severe white cast, the worst color reproduction, and less detail retention than the low-light image enhancement method proposed in this invention.

[0055] To verify that the low-light image enhancement method of this invention is effective on more low-light images, the Exdark dataset was fed into the low-light image enhancement network established in this invention to obtain enhanced images. Figure 3 This is a partial enhancement result of the method proposed in this invention on the Exdark dataset. From Figure 3The comparison images show that the existing MBLLEN method does not uniformly improve image brightness. In some images, it fails to separate the detected object contours from the darkness, allowing them to blend into the background. The ZeroDCE method significantly improves overall image brightness, but it amplifies image noise and introduces detrimental factors such as severe purple fringing. The KIND method also significantly improves image brightness, but... Figure 3 The noise in the image was amplified, and noticeable artifacts and white halos were produced. The method proposed in this invention performs well on the Exdark dataset, with minimal blurring of the original image, very few artifacts, and good enhancement effect on generally dark images without generating significant noise. However, the enhanced image exhibits a whitening effect.

[0056] This invention proposes a network structure for estimating brightness adjustment and noise mitigation components, which enhances image brightness while mitigating the impact of noise. The network for estimating the components is optimized using depthwise separable convolution and inverse residual structures. Furthermore, a network structure for estimating and optimizing the overall image's color correction matrix and gamma correction parameters is designed using scaling point attention, which effectively captures the overall image relationships, further optimizing the overall color and brightness of the image. Results on the LoL and Exdark datasets demonstrate that the method of this invention can improve the brightness of low-light images, preserve the details of low-light images, and suppress noise in low-light images.

Claims

1. A low-light image enhancement method based on parameter estimation, wherein the method uses a low-light image enhancement network to enhance the input low-light image; The low-light image enhancement network includes: Pixel-level component estimation branch and overall optimization branch; The pixel-level component estimation branch includes a first branch and a second branch with identical structures. The first or second branch, from input to output, sequentially includes: a first convolutional module, N inverse residual structures, and a second convolutional module. The input of the additional second convolutional module is a residual concatenation of the output of the first convolutional module and the outputs of the N inverse residual structures. The input of the first branch is the input of the low-light image enhancement network, and the input of the second branch is the result of a dot product between the output of the first branch and the input of the low-light image enhancement network, followed by concatenation with the input of the low-light image enhancement network. The output of the pixel-level component estimation branch is the sum of the output of the first branch and the output of the second branch. The overall optimization branch includes: a third convolutional module, a fourth convolutional module, a first fully connected layer, a second fully connected layer, a scaling attention module, a third fully connected layer, layer normalization, a fourth fully connected layer, an activation function, a fifth fully connected layer, a first convolutional layer, and a second convolutional layer. The input to the overall optimization branch is the concatenation of the input to the low-light image enhancement network and the output of the pixel-level component estimation branch. The input to the overall optimization branch passes through the third and fourth convolutional modules sequentially and is then output to the first and second fully connected layers, respectively. The input to the scaling attention module includes: the output of the first fully connected layer, the output of the second fully connected layer, and a query parameter. This query parameter is used to limit the output value of the overall optimization branch. The output of the scaling attention module sequentially passes through the third fully connected layer, layer normalization, the fourth fully connected layer, the activation function, and the fifth fully connected layer. The output of the third fully connected layer is residually concatenated with the output of the fifth fully connected layer and then added to a 3x3 identity matrix and a value of size 1. The result of the addition is split into a 3x3 matrix and a value of size 1. The 3x3 matrix is ​​input to the first fully connected layer, and the value of size 1 is input to the second fully connected layer. The output of the overall optimization branch is the result of matrix multiplication of the output of the pixel-level component estimation branch with the output of the first fully connected layer and then gamma transformation of the output of the second fully connected layer, which is the output of the low-light image enhancement network. The loss of the low-light image enhancement network as follows: ； ； ； in, The loss represents the pixel difference between the predicted image obtained by the low-light enhancement network and the corresponding high-brightness image in the LOL dataset. This represents the corresponding high-brightness image in the LOL dataset. The image represents the predicted image obtained by the low-light enhancement network. , , These represent the height, width, and number of channels of the real image, respectively, with superscript indicating the height, width, and number of channels. This represents the coordinates of the corresponding image feature in RGB space. , , Representing VGG-16 The height, width, and number of channels of the layer output features. This refers to the sequence number of the VGG-16 network layer. Representing the VGG-16 network The output of the layer, This represents the perceptual loss obtained using VGG-16.

2. The low-light image enhancement method based on parameter estimation as described in claim 1, characterized in that, The first and second convolutional modules in the pixel-level component estimation branch have the same structure, both including: a convolutional layer, a batch normalization layer, and an activation function in sequence; the inverse residual structure includes: a first pointwise convolutional module, a channelwise convolutional module, and a second pointwise convolutional module. The input of the first pointwise convolutional module is the input of the inverse residual structure, and the output of the inverse residual structure is the residual connection value between the output of the second pointwise convolutional module and the input of the inverse residual structure.

3. The low-light image enhancement method based on parameter estimation as described in claim 1, characterized in that, The number N of inverted residual structures in the first or second branch of the pixel-level component estimation branch is 6.

Images