A low-illumination image enhancement method based on semi-quadratic split optimization

Abstract

The application belongs to the technical field of digital image processing, and particularly relates to a low-illumination image enhancement method based on semi-quadratic splitting optimization; the method comprises the following steps: acquiring a low-illumination image and a normal-illumination image; inputting the low-illumination image into a decomposition network for processing to obtain a reflection map and an illumination map of the low-illumination image; inputting the reflection map and the illumination map of the low-illumination image into a deep unfolding network for processing to obtain an enhanced reflection map; inputting the normal-illumination image and the illumination map of the low-illumination image into an illumination enhancement network for processing to obtain an enhanced illumination map; and multiplying the enhanced illumination map and the enhanced reflection map pixel by pixel to obtain an enhanced low-illumination image; the application solves the problems of insufficient or excessive brightness enhancement, image color distortion and other degradation problems, and improves the performance of the network.

Description

Technical Field

[0001] This invention belongs to the field of digital image processing technology, specifically relating to a low-light image enhancement method based on semi-quadratic splitting optimization. Background Technology

[0002] With the continuous innovation and development of electronic communication technology, people have access to a plethora of channels for obtaining information. However, images, as one of the mediums for information transmission, still possess unparalleled advantages in terms of presentation format, access speed, and public acceptance. Therefore, the application of images is extremely widespread across various industries.

[0003] However, in real life, due to various objective reasons, including but not limited to weather conditions, hardware equipment, artificial light sources, shooting techniques, and economic levels, many images fail to achieve the desired effect. One such type is low-light images. Low-light images refer to images taken under conditions where natural or artificial light sources are too dim or unevenly lit. However, these types of images are widely used in our lives. In many cases, we urgently need to find valuable information from these degraded images. For example, in the defense field, special forces use night vision devices to observe objects in the dark during nighttime reconnaissance operations; in the aerospace field, low-light, low-contrast images captured by drones require image enhancement preprocessing before other processing techniques to obtain clearer images; in the industrial coal mining field, due to the limitations of the working environment, photos with low brightness and contrast and blurry image information are often produced, making it difficult to directly extract information from these images, thus affecting subsequent work; in the biomedical field, the grayscale distribution of medical images is determined by different parameters of human tissue characteristics. Generally, this contrast is very small, resulting in small differences in grayscale between adjacent images, making it difficult to accurately obtain information from the image visually. Therefore, image enhancement processing is necessary to provide doctors with more useful information.

[0004] Many solutions have emerged to address the problems of low-light image rendering. For example, algorithms based on grayscale transformation, while simple and convenient to operate, require manual parameter tuning and are prone to under- or over-enhancement. Histogram equalization algorithms can restore hidden details in dark areas, enhancing contrast and detail. However, if the histogram contains large peaks, the image will be over-enhanced, failing to achieve the desired effect. Furthermore, it does not consider noise. Algorithms based on Retinex theory can simultaneously consider image contrast, brightness, dynamic range, and color constancy, resulting in relatively balanced enhancement overall, but they also do not account for the impact of noise. Currently, deep learning-based algorithms are quite popular. These algorithms are data-driven and produce images with excellent visual effects, showing significant advantages in detail enhancement and image denoising. However, they also have drawbacks; deep learning network models require sufficient datasets to achieve good results, and the generalization ability of the network model is also a concern. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention proposes a low-light image enhancement method based on semi-quadratic splitting optimization. The method includes: acquiring the low-light image to be enhanced and inputting it into a trained low-light image enhancement model to obtain the enhanced image.

[0006] The training process for low-light image enhancement models includes:

[0007] S1: Acquire low-light and normal-light images; input the low-light image into the decomposition network for processing to obtain the reflectance and illumination maps of the low-light image;

[0008] S2: Input the reflectance map and illumination map of the low-light image into the depth unrolling network for processing to obtain the enhanced reflectance map;

[0009] S3: Input the illumination maps of the normal illumination image and the low illumination image into the illumination enhancement network for processing to obtain the enhanced illumination map;

[0010] S4: Multiply the enhanced illumination map and the enhanced reflectance map pixel by pixel to obtain the enhanced low-light image;

[0011] S5: Calculate the total model loss and adjust the model parameters based on the total model loss to obtain a trained low-light image enhancement model.

[0012] Preferably, the process by which the deep unrolling network processes the reflectance and illumination maps of low-light images includes:

[0013] The low-light image, along with its reflectance and illumination maps, are input into an implicit regularization model for processing to obtain intermediate enhanced reflectance parameters.

[0014] Set the number of iterations. If the current iteration does not meet the number of iterations, iteratively update the intermediate enhanced reflection parameters. Use the Model_R network to process the intermediate enhanced reflection parameters and the illumination map of the low-light image to obtain the enhanced reflection map, until the current iteration meets the number of iterations.

[0015] Furthermore, the implicit regularization model is expressed as:

[0016]

[0017] Where I represents the low-light image, L represents the illumination map of the low-light image, ° represents the product operation, R represents the reflectance map of the low-light image, and α represents the regularization weight. This represents the prior regularization term applied to the reflection graph R.

[0018] Furthermore, the intermediate enhanced reflection parameters are expressed as follows:

[0019]

[0020] Among them, P k R represents the intermediate enhanced reflectance parameters after the k-th iteration, β represents the penalty coefficient, R0 represents the initial low-light image reflectance map, and R k-1 Let I represent the reflectance map of the low-light image obtained after the (k-1)th iteration, where I represents the low-light image and L represents the illumination map of the low-light image.

[0021] Furthermore, the Model_R network includes 7 convolutional layers, 6 activation layers, and an SE attention module. Each of the first 6 convolutional layers is followed by an activation layer, which uses the ReLU activation function. The SE attention module is connected after the second convolutional layer.

[0022] Preferably, the process of obtaining the enhanced illumination map includes: the illumination enhancement network includes four convolutional activation layers and four attention extraction layers;

[0023] S31: Input the illumination map of the low-light image into the first convolutional activation layer to obtain the low-light feature map of the first layer;

[0024] S32: Input the normal illumination image into the current attention extraction layer to obtain the current attention feature map;

[0025] S33: Normalize the current attention feature map to obtain the intermediate normal illumination feature map; multiply the intermediate normal illumination feature map and the low illumination feature map of the current layer to obtain the low illumination feature map of the next layer;

[0026] S34: Use the low-light feature map of the next layer as the input of the next convolutional activation layer and return to step S32 until the fourth attention extraction layer outputs the attention feature map, and finally obtain the illumination map of the enhanced low-light image.

[0027] Preferably, the processing procedure of the attention extraction layer for the normal illumination image includes: performing three feature extractions on the normal illumination image to obtain three tensors A, B, and C; transforming the three tensors respectively to obtain tensors D, E, and F; performing matrix multiplication on tensor D and tensor E and then performing a softmax activation operation to obtain a weight matrix S; and multiplying the weight matrix S and tensor F and then summing it with the low illumination image I to obtain the attention feature map.

[0028] Preferably, the formula for calculating the total loss of the model is:

[0029] LOSS = loss decom +loss unfold +loss i_enhanced

[0030] Where loss represents the total loss of the model, loss decom The loss represents the total loss of the decomposed network. unfold The loss represents the total loss of the depth-unfolded network. i_enhanced This represents the total loss of the illumination-enhanced network.

[0031] The beneficial effects of this invention are as follows:

[0032] 1. Based on the Retinex theoretical model, this invention proposes a deep unfolding network based on a semi-quadratic split optimization algorithm, and uses a spatial position attention matrix in the illumination enhancement network, thereby obtaining higher quality and clearer reflection and illumination maps.

[0033] 2. The deep unfolded network proposed in this invention enhances the feature representation of reflection vectors, reduces the problem of blurred image details, and improves the generalization ability of the model.

[0034] 3. This invention uses spatial location attention to learn spatial feature correlation and contextual information, which solves the degradation problems such as insufficient or excessive brightness enhancement and image color distortion, thereby improving the network performance. Attached Figure Description

[0035] Figure 1 This is the low-light image enhancement model in this invention;

[0036] Figure 2 This is a schematic diagram of the attention feature map calculation process in this invention;

[0037] Figure 3 A comparison chart of image brightness in low-light conditions, using the URetinex method, and present invention.

[0038] Figure 4 Comparison of image details in low-light images, using the UREtinex method, and present invention. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] This invention proposes a low-light image enhancement method based on semi-quadratic splitting optimization, such as... Figure 1 As shown, the method includes the following: acquiring the low-light image to be enhanced and inputting it into a trained low-light image enhancement model to obtain the enhanced image.

[0041] The training process for low-light image enhancement models includes:

[0042] S1: Obtain low-light image and normal-light image; input the low-light image into the decomposition network for processing to obtain the reflectance map and illumination map of the low-light image.

[0043] Obtain training datasets of low-light and normal-light images; use a decomposition network to decompose the low-light and normal-light images.

[0044] The ability to decompose an image into a reflection image and an incident image is based on the Retinex theory proposed by the renowned physicist Land. This theory is derived by simulating the imaging principles of the retina and cortex in the human brain. Specifically, it states that the amount of visible light reaching the human eye can be expressed as the product of reflectivity and illumination component.

[0045] S(x,y)=L(x,y)°R(x,y)

[0046] S(x,y) represents the original image, L(x,y) represents the incident component or illumination intensity, R(x,y) represents the reflected component, and ° represents the product operation.

[0047] The decomposition network consists of 5 convolutional layers and 5 activation layers, with each convolutional layer followed by an activation layer. The first four activation layers use the LeakyReLU activation function, and the last activation layer uses the ReLU activation function. The specific parameters of this decomposition network are shown in Table 1. The first convolutional layer has 3 input channels and 32 output channels, the middle three layers each have 32 input and 32 output channels, and the last layer has 32 input channels and 4 output channels. The number of channels represents the dimensionality of information contained in the model output. Generally, the more channels, the stronger the ability to represent information. Since the reflection layer contains more detailed components, the feature vectors of the first three channels are taken as the reflection component of the reflection map, and the feature vector of the last channel is taken as the illumination component of the illumination map. To solve the gradient vanishing problem, the LeakyReLU activation function is added after each convolutional layer. By introducing a small slope in the negative value region, the gradient vanishing problem is alleviated, and the generalization ability of the model can also be improved.

[0048] Table 1. Specific parameters of the decomposition network

[0049] Conv1+Leaky 3 32 3*3 1 1 Conv2+Leaky 32 32 3*3 1 1 Conv3+Leaky 32 32 3*3 1 1 Conv4+Leaky 32 32 3*3 1 1 Conv5+ReLU 32 4 3*3 1 1

[0050] S2: Input the reflectance map and illumination map of the low-light image into the depth unrolling network for processing to obtain the enhanced reflectance map.

[0051] The process of using a deep unrolled network to process the reflectance map of a low-light image:

[0052] The low-light image, along with its reflectance and illumination maps, are input into an implicit regularization model for processing to obtain the intermediate enhanced reflectance parameter P; specifically:

[0053] Based on the Retinex theoretical model, this invention uses a semi-quadratic splitting method to transform the image decomposition problem formula into an implicit regularization model; because image decomposition is an ill-posed solution problem, a regularization (prior) term is used to restrict the solution space. This implicit regularization model is expressed as:

[0054]

[0055] Where I represents the low-light image, and L represents the illumination map of the low-light image. This represents a product operation, where R represents the reflectance of the low-light image. This indicates that the calculation is performed using the F-norm, and α represents the trainable regularization weight parameter. This represents the prior regularization term applied to the reflection map R. The smaller the value of the function, the closer the product of the reflection map and the illumination map obtained from the decomposition is to the original map.

[0056] The first term of the function is the data fidelity term, which ensures that the result conforms to the noise reduction process. The second term is the prior term, which enhances the output. To facilitate optimization, a semi-quadratic split is used, and a variable P (intermediate enhancement reflection parameter) is inserted, resulting in the following constrained optimization equation:

[0057]

[0058] To transform a constrained optimization problem into an unconstrained optimization problem, a quadratic penalty term can be introduced, allowing direct control over the penalty term to obtain the solution to the optimization problem.

[0059]

[0060] Where β represents the penalty coefficient, R k-1 Let I represent the low-light image reflectance map obtained after the (k-1)th iteration, where I represents the low-light image.

[0061] Based on the principle of alternating updates for one variable and keeping the other variable constant, the sub-variables P and R are iteratively solved. The expressions for sub-variables P and R are shown below:

[0062]

[0063]

[0064] Here, k represents the current iteration number, which can be manually set using the controlled variable method. With other conditions remaining constant, changing the iteration number and comparing the evaluation metrics of the final augmented graph, along with the values ​​of the evaluation metrics, the number of iterations, and the model's efficiency, determines the appropriate number of iterations.

[0065] Solving for the subvariable P: We can set its first derivative to zero, and then we get the following expression:

[0066]

[0067] Among them, P k R represents the intermediate enhanced reflectance parameters after the k-th iteration; R0 represents the initial low-light image reflectance map, and the value of P1 is R0; R k-1 This represents the low-light image reflectance map obtained after the (k-1)th iteration.

[0068] The iteration count is set. If the current iteration does not meet the required number of iterations, the intermediate enhanced reflection parameters are updated iteratively. The Model_R network is used to process the intermediate enhanced reflection parameters and the illumination map of the low-light image to obtain the enhanced reflection map. This process continues until the current iteration meets the required number of iterations. Specifically:

[0069] Set the iteration number k to control the number of iterations, and each iteration contains two phases.

[0070] When k=1, the first stage is: updating the value of the sub-variable P (the initial value of the sub-variable P is the reflectance map obtained by decomposing the low-light image). The second stage is: model training. The reflectance map P1 and the illumination map L of the low-light image are input into the Model_R network. After training, the reflectance tensor R1 is obtained.

[0071] When k=2, the first stage involves updating the value of the sub-variable P. The low-light image I, the reflection tensor R1 from the previous iteration, the illumination map L, and the trainable parameter weights β1 (initially randomly set) are input into the formula for updating the sub-variable P, resulting in tensor P2. The second stage involves model training. Tensor P2 and the illumination map L are input into the Model_R network, and after training, the reflection tensor R2 is obtained.

[0072] Similarly, when k = n (where n is the maximum number of iterations), the first stage involves updating the value of the sub-variable P. This involves updating the low-light image I and the reflection tensor R from the previous iteration. n-1 Illumination map L and trainable parameter weights β n-1 The four values ​​are input into the formula for updating the sub-variable P to obtain the tensor P. n The second stage: Model training. Tensor P... n The illumination map L is input into the Model_R network, and after training, the reflection tensor R is obtained. n R n That is, the final enhanced reflection map.

[0073] After the Model_R network receives the updated sub-variable P and the illumination map L of the low-light image as input, it first concatenates them in the second dimension to obtain tensor x. Then, tensor x is input into the conv1 convolutional layer to obtain x1, and x1 is input into the SE attention module to obtain tensor x2. Next, x2 is input into the conv2 convolutional layer to obtain x3, and so on, repeating four times. Finally, tensor x7 is input into the conv7 convolutional layer to obtain tensor n. Tensor n and the input variable P are summed to obtain the reflection tensor R. This invention adds an SE attention module after the conv1 convolutional layer to increase the extraction of useful information and suppress interference from useless information. The Model_R network consists of 7 convolutional layers, 6 activation layers, and an SE attention module. Each of the first 6 convolutional layers is followed by an activation layer using the ReLU activation function. The SE attention module is connected after the second convolutional layer. The specific parameters of the Model_R network are shown in Table 2.

[0074] Table 2 Specific parameters of Model_R

[0075]

[0076]

[0077] The initial values ​​for a deep unfolded network are the illumination and reflection maps obtained by decomposing the network, rather than random matrices or all-zero and all-one matrices. This greatly improves the generalization ability of the network model and the ability to enhance the details of the enhanced reflection map.

[0078] S3: Input the illumination map of the normal illumination image and the illumination map of the low illumination image into the illumination enhancement network for processing to obtain the enhanced illumination map.

[0079] The illumination enhancement network consists of four convolutional activation layers and four attention extraction layers. To obtain a larger receptive field and capture broader contextual information, the kernel size is set to 5*5. Simultaneously, to maintain the consistent width and height of the feature vectors, the stride is set to 1 and the padding to 2.

[0080] The process of processing the low-light illumination map obtained by the illumination enhancement network includes:

[0081] The low-light image illumination map L obtained from the decomposition is input into the first convolutional activation layer to obtain the feature map conv1 of the first layer; at the same time, the normal illumination image I is input into the first convolutional activation layer to obtain the feature map conv1 of the first layer; * The input is fed into the first attention extraction layer to obtain the first attention feature map atmap1. Atmap1 is then normalized using softmax to obtain map1. map1 is multiplied by the corresponding feature map conv1 to obtain conv2. Similarly, conv2 is used as the input to the second convolutional activation layer. After passing through the second attention extraction layer and normalization, map2 is obtained. Map2 is multiplied by the second feature map conv2 to obtain the input to the third convolutional activation layer. This process of convolutional activation layer, attention extraction layer, and softmax normalization is repeated twice more to finally obtain the illumination map of the enhanced low-light image.

[0082] The processing of normal illumination images by the attention extraction layer includes:

[0083] like Figure 2 As shown, normal illumination image I * Tensors A, B, and C were obtained after three feature extractions, and The three tensors are transformed separately: tensor A is reshaped and transposed to obtain tensor D; tensor B is reshaped to obtain tensor E; and tensor C is reshaped to obtain tensor F. Multiplying tensor D and tensor E by matrix multiplication followed by softmax activation yields a weight matrix S. The product of weight matrix S and tensor F is then summed with the low-light image I to obtain the final result (attention feature map). This process constitutes the entire attention tensor extraction process.

[0084] S4: Multiply the enhanced illumination map and the enhanced reflectance map pixel by pixel to obtain the enhanced low-light image.

[0085] S5: Calculate the total model loss and adjust the model parameters based on the total model loss to obtain a trained low-light image enhancement model.

[0086] The decomposed total loss function of the network is:

[0087] loss decom =loss rec +ω rs loss rs +ω is loss is

[0088]

[0089] loss rs =||RR * ||1

[0090]

[0091] Where, loss decom This represents the total loss of the decomposition network, including image restoration loss. rec Reflection consistency loss rs and illumination smoothness loss is Three parts. I, R, and L represent the original image, reflectance image, and illumination image of the low-light image, respectively; I * R * L * These represent the original image, reflectance image, and illumination image under normal illumination, respectively. This indicates calculating the gradient of an image, including the horizontal gradient. and the gradient in the vertical direction ° indicates the product of elements; ||·||1 indicates the use of the L1 norm for calculation; ε is set to 0.01 to prevent the denominator from being zero; ω rs ω is These are reflection consistency loss. rsand illumination smoothness loss is The weighting coefficients are 0.01 and 0.15.

[0092] The total loss function of the deep unfolded network is:

[0093] loss unfold =loss rf +loss re

[0094]

[0095]

[0096] Where, loss unfold This represents the total loss of the depth-unfolded network, including the reflection fidelity term. rf and reflection gradient term loss fg It consists of two parts. SSIM(·) represents the structural similarity index. The larger the structural similarity index value, the more similar the two input images are, which contradicts the minimization of the loss function, so it is taken as a negative number. Represents the reflection map during training; R * A reflectance diagram representing normal illuminance; This indicates calculating the gradient of an image, including the horizontal gradient. and the gradient in the vertical direction ‖·‖1 indicates that the L1 norm is used for calculation.

[0097] The total loss function of the illumination enhancement network is:

[0098] loss i_enhanced =loss if +loss im

[0099]

[0100]

[0101] Where, loss u_enhanced This represents the total loss of the illumination enhancement network, including illumination consistency loss. if and exposure consistency loss im . Represents the illumination map during the training process; L * L represents the illumination diagram of a normal illumination image; L represents the illumination diagram of a low illumination image. This indicates calculating the gradient of an image, including the horizontal gradient. and the gradient in the vertical direction c control curve The steepness of the slope, where c = 10.

[0102] Therefore, the total loss of the model is:

[0103] LOSS = loss decom +loss unfold +loss i_enhanced

[0104] The low-light image to be enhanced is obtained and then input into a trained low-light image enhancement model to obtain the enhanced image.

[0105] Evaluation of the present invention: such as Figure 3 As shown, the first column represents the selected low-light image, the second column is the result image after passing through the URetinex network, and the third column is the effect image obtained after adaptively adjusting the illumination according to the present invention. It can be seen from the above images that the URetinex-enhanced image shows insufficient enhancement for images with very low illumination; in Figure 4 In comparison between images processed by the UREtinex network and images from the present invention, it was found that the present invention significantly improves image distortion, contrast, and brightness.

[0106] The above-described embodiments further illustrate the purpose, technical solution, and advantages of the present invention. It should be understood that the above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made to the present invention within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A low-light image enhancement method based on semi-quadratic splitting optimization, characterized in that, include: The low-light image to be enhanced is obtained and input into the trained low-light image enhancement model to obtain the enhanced image. The training process for low-light image enhancement models includes: S1: Acquire low-light and normal-light images; input the low-light image into the decomposition network for processing to obtain the reflectance and illumination maps of the low-light image; S2: Input the reflectance map and illumination map of the low-light image into the depth unrolling network for processing to obtain the enhanced reflectance map; S3: Input the illumination maps of the normal illumination image and the low illumination image into the illumination enhancement network for processing to obtain the enhanced illumination map; S4: Multiply the enhanced illumination map and the enhanced reflectance map pixel by pixel to obtain the enhanced low-light image; S5: Calculate the total model loss and adjust the model parameters based on the total model loss to obtain a trained low-light image enhancement model.

2. The low-light image enhancement method based on semi-quadratic splitting optimization according to claim 1, characterized in that, The decomposition network consists of 5 convolutional layers and 5 activation layers, with each convolutional layer followed by an activation layer. The first 4 activation layers use the LeakyReLU activation function, and the last activation layer uses the ReLU activation function.

3. The low-light image enhancement method based on semi-quadratic splitting optimization according to claim 1, characterized in that, The process by which a deep unfolding network processes the reflectance and illumination maps of a low-light image includes: The low-light image, along with its reflectance and illumination maps, are input into an implicit regularization model for processing to obtain intermediate enhanced reflectance parameters. Set the number of iterations. If the current iteration does not meet the number of iterations, iteratively update the intermediate enhanced reflection parameters. Use the Model_R network to process the intermediate enhanced reflection parameters and the illumination map of the low-light image to obtain the enhanced reflection map, until the current iteration meets the number of iterations.

4. The low-light image enhancement method based on semi-quadratic splitting optimization according to claim 3, characterized in that, The implicit regularization model is expressed as: Where I represents the low-light image, L represents the illumination map of the low-light image, ° represents the product operation, R represents the reflectance map of the low-light image, and α represents the regularization weight. This represents the prior regularization term applied to the reflection graph R.

5. The low-light image enhancement method based on semi-quadratic splitting optimization according to claim 3, characterized in that, The intermediate enhanced reflection parameters are expressed as follows: Among them, P k R represents the intermediate enhanced reflectance parameters after the k-th iteration, β represents the penalty coefficient, R0 represents the initial low-light image reflectance map, and R k-1 Let I represent the reflectance map of the low-light image obtained after the (k-1)th iteration, and let L represent the illumination map of the low-light image.

6. The low-light image enhancement method based on semi-quadratic splitting optimization according to claim 3, characterized in that, The Model_R network consists of 7 convolutional layers, 6 activation layers, and an SE attention module. Each of the first 6 convolutional layers is followed by an activation layer, which uses the ReLU activation function. The SE attention module is connected after the second convolutional layer.

7. The low-light image enhancement method based on semi-quadratic splitting optimization according to claim 1, characterized in that, The process of obtaining the enhanced illumination map includes: the illumination enhancement network consists of four convolutional activation layers and four attention extraction layers; S31: Input the illumination map of the low-light image into the first convolutional activation layer to obtain the low-light feature map of the first layer; S32: Input the normal illumination image into the current attention extraction layer to obtain the current attention feature map; S33: Normalize the current attention feature map to obtain the intermediate normal illumination feature map; multiply the intermediate normal illumination feature map and the low illumination feature map of the current layer to obtain the low illumination feature map of the next layer; S34: Use the low-light feature map of the next layer as the input of the next convolutional activation layer and return to step S32 until the fourth attention extraction layer outputs the attention feature map, and finally obtain the illumination map of the enhanced low-light image.

8. The low-light image enhancement method based on semi-quadratic splitting optimization according to claim 1, characterized in that, The attention extraction layer's processing of the normal illumination image includes: performing three feature extractions on the normal illumination image to obtain three tensors A, B, and C; transforming the three tensors to obtain tensors D, E, and F; performing matrix multiplication on tensors D and E followed by softmax activation to obtain a weight matrix S; and multiplying the weight matrix S and tensor F together with the low illumination image I to obtain the attention feature map.

9. The low-light image enhancement method based on semi-quadratic splitting optimization according to claim 1, characterized in that, The formula for calculating the total loss of the model is: LOSS=loss decom +loss unfold +loss i_enhance Where loss represents the total loss of the model, loss decom The loss represents the total loss of the decomposed network. unfold The loss represents the total loss of the depth-unfolded network. i_enhanced This represents the total loss of the illumination-enhanced network.

Images