Image enhancement method and system in mine complex limited space low-illumination environment

Abstract

The invention discloses an image enhancement method and system in a mine complex limited space low-illumination environment, and the method comprises the steps: carrying out the contrast preprocessing of a low-light image through an image preprocessing module, carrying out the extraction of the preprocessed image through a low-frequency information extraction network, obtaining low-frequency information, subtracting the low-frequency information from an input image, and obtaining high-frequency information; and fusing the low-frequency information and the high-frequency information through a feature fusion network, inputting the fused information into an image enhancement network for enhancement, and inputting the enhanced image into the pre-trained LPDM diffusion model for post-processing to obtain a final image. The system comprises a camera, an image preprocessing module, a frequency processing and fusing module and a post-processing and optimizing module. According to the method and the system, the quality of the low-light image is effectively improved, the recognizable degree of the image is improved, and the efficiency and the accuracy of image processing in the low-light environment are improved.

Description

Technical Field

[0001] The invention relates to an image enhancement method and system in a low-illumination environment of a complex confined space in a mine, and belongs to the technical field of image processing. Background Art

[0002] In the special scenario of underground coal mine operations, many factors such as the closed space, the limitation of light sources, the complexity of the mine tunnel structure, and the presence of dust and smoke make it difficult for artificial light sources to achieve uniform coverage during illumination, thus forming an environment of uneven lighting and low illumination. This situation has a great negative impact on the visual monitoring system in coal mines, resulting in poor quality of collected images and blurred details, which seriously interferes with the safe operation process of miners, impacts the timeliness and accuracy of accident warnings, and also hinders the efficient implementation of emergency response work.

[0003] In such low-light environments, traditional image processing methods are difficult to effectively extract useful information from images, which has a substantial impact on the safety management and real-time monitoring of mines. However, from the perspective of the frequency characteristics of the image itself, in-depth analysis shows that low-light images of mines also contain rich information, including high-frequency information and low-frequency information. High-frequency information carries key details such as the texture of the tunnel wall and the fine lines on the surface of the equipment. These details are crucial for accurately identifying specific features in the underground environment; low-frequency information reflects the overall structural conditions such as the overall outline of the tunnel and the layout of large equipment, providing an important basis for us to grasp the overall layout of the underground environment.

[0004] However, most of the current low-light image enhancement schemes do not make detailed distinctions between low-light images under different light intensities. When processing brighter low-light scenes, over-enhancement may lead to overexposure. When faced with extremely dark low-light images, it is difficult to fully tap their potential detail information and cannot effectively improve the image recognizability. Summary of the invention

[0005] The purpose of the invention is to provide an image enhancement method and system in a low-light environment in a complex confined space of a mine. The method and system can effectively improve the quality of low-light images, enhance the recognizability of images, and improve the efficiency and accuracy of image processing in a low-light environment.

[0006] In order to achieve the above object, the present invention provides an image enhancement method in a complex confined space of a mine under low illumination environment, comprising the following steps:

[0007] S1, performing contrast preprocessing on the low-light image through an image preprocessing module;

[0008] S2. Extract the low-frequency information from the preprocessed image through a low-frequency information extraction network, and then subtract the low-frequency information from the input image to obtain the high-frequency information;

[0009] S3. Fuse the low-frequency information and the high-frequency information through a feature fusion network and then input them into an image enhancement network for enhancement;

[0010] S4. Input the enhanced image into the pre-trained LPDM diffusion model for post-processing to obtain the final image.

[0011] Further, the specific process of image preprocessing in S1 is as follows:

[0012] S1.1. Adjust the low-light image to be processed to the size of the set pixels;

[0013] S1.2. Divide the low-light image into multiple regions, and perform histogram equalization on each region respectively. The specific steps are as follows:

[0014] S1.2.1. Divide the image into K×L sub-blocks B of size m×n ij , where i = 0, 1,... K-1, j = 0, 1,..., L-1;

[0015] S1.2.2. For each sub-block B ij , calculate its gray histogram H ij (k), where k = 0, 1,... G-1, and G is the number of gray levels;

[0016] S1.2.3. Calculate the cumulative distribution function C ij of each sub-block B ij . The formula is:

[0017] S1.2.4. For each pixel I(x, y) in sub-block B ij , its new gray value where represents the floor operation;

[0018] S1.3. After inputting the low-light image i1, first perform adaptive histogram equalization to obtain i2. Then, based on the image i2, adjust it within the contrast adjustment coefficient range of 0.6 - 0.8 to obtain i3. Next, continue to adjust the contrast of i3 and select within the contrast adjustment coefficient range of 0.3 - 0.5 to obtain i4. Finally, connect i1, i2, i3, and i4. During this period, according to the brightness and detail changes in each region of the image, ensure that different contrast images contain feature information under different illuminations. Then connect these four images at the channel layer to obtain I(x, y). The connection preserves the original information of each feature map, enabling the subsequent network layers to make full use of this information and avoiding information mixing and loss. At this time, the dimension information is 12 * 256 * 256, completing the preprocessing of the image.

[0019] Further, the specific process of S2 is as follows:

[0020] S2.1. Use a convolutional mean filter to extract the low-frequency information L(x, y) of the upper image I(x, y), and then subtract the low-frequency information from I(x, y) to obtain the high-frequency information.

[0021] Specifically, let the convolutional kernel be K with a size of m×n. The formula for calculating the low-frequency information is: The formula for calculating the high-frequency information is:

[0022] H(x, y) = I(x, y) - L(x, y);

[0023] S2.2. Use a dual-branch frequency enhancement module to enhance the high-frequency and low-frequency information extracted in S2.1. Each path uses a different number of convolutions. One branch is equipped with three convolutional layers, and a channel attention module is added before convolution. At the same time, use a residual connection to add the input at the beginning to the output after the third convolution, which focuses more on the high-frequency details of the image. The other branch uses dilated convolution, which pays more attention to the low-frequency content and contours of the image. Finally, connect the results of these two branches. Among them, the formula for the high-frequency branch module is:

[0024] F high = F + I;

[0025] F = Conv(Conv(Conv(CA(I))));

[0026] The formula for the low-frequency branch module is:

[0027] F low = DConv(I);

[0028] Among them, I is the input feature map, Conv(.) is the convolution operation, CA(.) represents the channel attention module, DConv(.) represents the dilated convolution operation, F represents the feature passing through the convolution and attention modules, and F high represents the enhanced high-frequency information obtained.

[0029] Furthermore, the specific process of S3 is as follows: The enhanced high-frequency and low-frequency information are concatenated at the channel, and the enhanced image is obtained through the feature enhancement module. The feature enhancement module adds DenseNet between the first network and the second network and adds the color enhancement attention module;

[0030] The specific process is as follows:

[0031] S3.1. Convolution layer processing: F conv = ReLU(Conv(I, w conv ));

[0032] S3.2. DenseNet processing: F dense = DenseNet(F conv );

[0033] S3.3. Subsequent convolution layer and activation function processing:

[0034] The first convolution layer: F 1 = ReLU(F dense , w 1 );

[0035] The second convolution layer: F 2 = ReLU(F 1 , w 2 );

[0036] The third convolution layer: F 3 = Tanh(CEM(F 2 ), w 3 );

[0037] Among them, ReLU and Tanh are activation functions, DenseNet(.) is the function of DenseNet, and CEM is the color enhancement attention module.

[0038] Furthermore, the training of the LPDM diffusion model in S4 is a noise detector, which is divided into two stages: training and inference; the specific processes are as follows:

[0039] S4.1. Training stage: Sample noise ∈ from the standard normal distribution, and at the same time obtain the initial real image samples x 0 in the mine scene, including mine roadway images and equipment images. At time step t, according to the probability distribution q(x t|x 0 ) Add noise to x 0 to obtain noisy mine image data x t , combine with the condition information c related to the mine, add the two element-wise and then input into the diffusion model; the diffusion model outputs the predicted noise ∈′, and the loss function L DM (∈, ∈′) is used to train the model to improve its accuracy in predicting mine image noise; the loss function L DM has the following calculation formula:

[0040]

[0041] where, θ is the parameter basis for the diffusion model to process the input; ∈ θ is the diffusion model function based on the parameter θ; α t is the key variable α defined to simplify the derivation and expression of the diffusion model-related formulas t = 1 - β t where β t ∈ (0, 1) is the variance parameter for controlling the noise addition in the forward diffusion process at time step t;

[0042] S4.2. Inference stage: The condition information c related to the mine is first processed by the LLIE backbone network to obtain This result is added element-wise with the relevant quantities and then input into the diffusion model; the diffusion model processes the input according to the parameter θ and outputs and then the denoised mine image is obtained The function is further processed based on and the parameter s, where The generated result is compared with the real image x 0 for comparative evaluation:

[0043]

[0044] where, α s is the variable derived from the variance parameter for controlling the noise addition in the forward diffusion process;

[0045] S4.3. To make the enhanced image more in line with human vision and improve the generalization ability of the model so that it can achieve better enhancement effects on different low-light image data, a custom loss function is used, L ssim is the structural similarity loss, L col is the color constancy loss:

[0046] L ssim = 1 - SSIM(x, y);

[0047]

[0048] Among them, x and y are two images to be measured, and μ x represents the average value of x pixels, and μ y represents the average value of y pixels, σ x and σ y are the variances of the images, σ xy is the covariance of x and y, C 1 and C 2 are two constants used to maintain stability;

[0049] Since L ssim does not consider the constraint of color information, the color constancy loss L col is used to correct color distortion:

[0050]

[0051] Among them, ρ = {(R, G), (R, B), (G, B)}, I p is the average channel value of channel P in the enhanced image; I q is the average channel value of channel q in the reference image.

[0052] Data acquisition adopts contrastive learning:

[0053]

[0054] Among them, L i represents the contrastive learning loss of the i-th sample, z i is the feature representation of the current sample, z i + is the feature representation of the positive sample corresponding to z i , N is the total number of samples including positive and negative samples, z j represents the feature representation of the j-th sample, sim(.,.) is the similarity function, and τ is the temperature hyperparameter used to adjust the difficulty and sensitivity of contrastive learning.

[0055] An image enhancement system for low-light environments in complex confined spaces in mines, including a camera, an image preprocessing module, a frequency processing and fusion module, and a postprocessing and optimization module;

[0056] The described camera is installed at the intersection of roadways and at the equipment operation points, and is used to collect the original image data in the low-light environment of the mine and transmit the image data to the image preprocessing module in real time;

[0057] The described image preprocessing module receives the image data transmitted by the camera, internally integrates an adaptive histogram equalization module and a contrast adjustment module to form a multi-contrast input image, and then outputs it to the frequency processing and fusion module;

[0058] The described frequency processing and fusion module uses a convolutional mean filter to separate the high and low frequency information of the preprocessed image; the dual-branch processing module strengthens the information according to the frequency characteristics. The high-frequency branch highlights details through three layers of convolution, a channel attention module, and residual connections, while the low-frequency branch uses one layer of convolution to focus on the low-frequency content; the frequency fusion module combines a specific convolutional layer and DenseNet layer to fuse and enhance the high and low frequency information, and outputs it to the post-processing and optimization module;

[0059] The described post-processing and optimization module uses the diffusion model LPDM to remove image noise, and optimizes and adjusts through a custom loss function that includes structural similarity, color constancy, and contrast learning loss, corrects colors, preserves structures, and enhances feature representations, and outputs high-quality images that meet the requirements of mine operations.

[0060] In the present invention, the image preprocessing module performs contrast preprocessing on the low-light image, extracts the low-frequency information from the preprocessed image through the low-frequency information extraction network, then subtracts the low-frequency information from the input image to obtain the high-frequency information, fuses the low-frequency information and the high-frequency information through the feature fusion network and inputs it into the image enhancement network for enhancement, and inputs the enhanced image into the pre-trained LPDM diffusion model for post-processing to obtain the final image. The low-light image enhancement method of the present invention realizes more comprehensive, effective, and high-quality low-light image enhancement through multi-contrast input construction, frequency separation and dual-branch processing, frequency fusion and contrast learning, and post-processing and detail retention operations, effectively improves the quality of low-light images, enhances the recognizability of images, and improves the efficiency and accuracy of image processing in low-light environments. Description of the Drawings

[0061] Figure 1 is a schematic diagram of the working process of the method of the present invention;

[0062] Figure 2 is a schematic diagram of the image processing process of the present invention;

[0063] Figure 3 is a schematic diagram of the dual-branch frequency enhancement module of the present invention;

[0064] Figure 4 is a schematic diagram of the feature fusion network of the present invention;

[0065] Figure 5 is a schematic diagram of the training stage of the LPDM diffusion model of the present invention;

[0066] Figure 6 is a schematic diagram of the inference stage of the LPDM diffusion model of the present invention;

[0067] Figure 7(a) is a low-light image, and (b) is the enhanced image processed by the present invention. Detailed implementation manners

[0068] The present invention will be further described below with reference to the accompanying drawings.

[0069] As Figure 1 shown, an image enhancement method for a complex and restricted space with low illuminance in a mine includes the following steps:

[0070] S1. Preprocess the contrast of the low-light image through an image preprocessing module;

[0071] S2. Extract the low-frequency information from the preprocessed image through a low-frequency information extraction network, and then subtract the low-frequency information from the input image to obtain the high-frequency information;

[0072] S3. Fuse the low-frequency information and the high-frequency information through a feature fusion network and input the fused information into an image enhancement network for enhancement;

[0073] S4. Input the enhanced image into the pre-trained LPDM diffusion model for post-processing to obtain the final image.

[0074] Further, as Figure 2 shown, the specific process of the image preprocessing in S1 is as follows:

[0075] S1.1. Adjust the size of the low-light image to be processed to the set size of 256*256 pixels;

[0076] S1.2. Divide the low-light image into multiple regions, and perform histogram equalization on each region respectively. The specific steps are as follows:

[0077] S1.2.1. Divide the image into K×L sub-blocks B of size m×n ij , where i = 0, 1,... K-1, j = 0, 1,..., L-1;

[0078] S1.2.2. For each sub-block B ij , calculate its gray-level histogram H ij (k), where k = 0, 1,... G-1, and G is the number of gray levels;

[0079] S1.2.3. Calculate the cumulative distribution function C ij of each sub-block B ij . The formula is:

[0080] S1.2.4. For each pixel I(x, y) in sub-block B ij , its new gray value wherein represents the floor operation;

[0081] S1.3. After inputting the low-light image i1, first perform adaptive histogram equalization to obtain i2. Then, based on the image i2, adjust it within the contrast adjustment coefficient range of 0.6 - 0.8 to obtain i3. Next, continue to adjust the contrast of i3 and select within the contrast adjustment coefficient range of 0.3 - 0.5 to obtain i4. Finally, connect i1, i2, i3, and i4. During this period, according to the brightness and detail changes in each region of the image, ensure that the images with different contrasts contain the characteristic information under different illuminations. Then, connect these four images at the channel layer to obtain I(x, y). The connection preserves the original information of each feature map, enabling the subsequent network layers to make full use of this information, avoiding information mixing and loss. At this time, the dimension information is 12 * 256 * 256, completing the preprocessing of the image.

[0082] Furthermore, the specific process of S2 is as follows:

[0083] S2.1. Use a convolutional mean filter to extract the low-frequency information L(x, y) of the upper image I(x, y), and then subtract the low-frequency information from I(x, y) to obtain the high-frequency information;

[0084] Specifically, let the convolutional kernel be K with a size of m×n. The calculation formula for the low-frequency information is: The calculation formula for the high-frequency information is:

[0085] H(x, y) = I(x, y) - L(x, y);

[0086] S2.2. As Figure 3 shown, use a dual-branch frequency enhancement module to enhance the high-frequency and low-frequency information extracted in S2.1. Each path uses a different number of convolutions. One branch is equipped with three convolutional layers, and a channel attention module is added before convolution. At the same time, use a residual connection to add the input at the beginning to the output after the third convolution, which focuses more on the high-frequency details of the image; the other branch uses dilated convolution, which pays more attention to the low-frequency content and contours of the image. Finally, connect the results of these two branches; among them, the formula for the high-frequency branch module is:

[0087] F high = F + I;

[0088] F = Conv(Conv(Conv(CA(I))));

[0089] The formula for the low-frequency branch module is:

[0090] F low = DConv(I);

[0091] Wherein, I is the input feature map, Conv(.) is the convolution operation, CA(.) represents the channel attention module, DConv(.) represents the dilated convolution operation, F represents the feature passing through the convolution and attention modules, and F high represents the enhanced high-frequency information obtained.

[0092] Furthermore, as Figure 4 shown, the specific process of S3 is as follows: Concatenate the enhanced high-frequency and low-frequency information at the channel, and obtain the enhanced image through the feature enhancement module. The feature enhancement module adds DenseNet between the first network and the second network, and adds a color enhancement attention module;

[0093] The specific process is as follows:

[0094] S3.1. Convolution layer processing: F conv = ReLU(Conv(I, w conv ));

[0095] S3.2. DenseNet processing: F dense = DenseNet(F conv );

[0096] S3.3. Subsequent convolution layer and activation function processing:

[0097] The first convolution layer: F 1 = ReLU(F dense , w 1 );

[0098] The second convolution layer: F 2 = ReLU(F 1 , w 2 );

[0099] The third convolution layer: F 3 = Tanh(CEM(F 2 ), w 3 );

[0100] Wherein, ReLU and Tanh are activation functions, DenseNet(.) is the function of DenseNet, and CEM is the color enhancement attention module.

[0101] Furthermore, the training of the LPDM diffusion model in S4 is a noise detector, which is divided into two stages: training and inference; the specific processes are as follows:

[0102] S4.1. Training stage: As Figure 5 shown, sample noise ∈ from the standard normal distribution, and at the same time obtain the initial real image sample x under the mine scene0 , including mine roadway images and equipment images. At time step t, according to the probability distribution q(x t |x 0 ), noise is added to x 0 to obtain noisy mine image data x t . Combining the conditional information c related to the mine, after adding them element by element, the result is input into the diffusion model; the diffusion model outputs the predicted noise ∈′, and the loss function L DM (∈, ∈′) is calculated to train the model to improve its accuracy in predicting the noise of mine images; the calculation formula of the loss function L DM is as follows:

[0103]

[0104] where θ is the parameter basis for the diffusion model to process the input; ∈ θ is the diffusion model function based on the parameter θ; α t is a key variable defined to simplify the derivation and expression of the formulas related to the diffusion model, α t = 1 - β t where β t ∈ (0, 1) is the variance parameter that controls the noise addition in the forward diffusion process at time step t;

[0105] S4.2. Inference stage: As Figure 6 shown, the conditional information c related to the mine is first processed by the LLIE backbone network to obtain x 0 μ . After adding this result element by element with the relevant quantities, it is input into the diffusion model; the diffusion model processes the input according to the parameter θ and outputs and then obtains the denoised mine image function is further processed based on x 0 μ , and the parameter s, where compares the generated result with the real image x 0 for evaluation:

[0106]

[0107] where α s is a variable derived from the variance parameter used to control the noise addition in the forward diffusion process;

[0108] S4.3. In order to make the enhanced image more in line with human vision and improve the generalization ability of the model so that it can achieve better enhancement effects on different low-light image data, a custom loss function is used. L ssim is the structural similarity loss, Lcol is the color constancy loss:

[0109] L ssim = 1 - SSIM(x, y);

[0110]

[0111] where x and y are two images to be measured, μ x represents the average value of x pixels, μ y represents the average value of y pixels, σ x and σ y are the variances of the images, σ xy is the covariance of x and y, C 1 and C 2 are two constants used to maintain stability;

[0112] Since L ssim does not consider the constraints of color information, the color constancy loss L col is used to correct color distortion:

[0113]

[0114] where ρ = {(R, G), (R, B), (G, B)}, I p is the average channel value of channel P in the enhanced image; I q is the average channel value of channel q in the reference image;

[0115] Data acquisition uses contrastive learning:

[0116]

[0117] where L i represents the contrastive learning loss of the i-th sample, z i is the feature representation of the current sample, z i + is the feature representation of the positive sample corresponding to z i , N is the total number of samples including positive and negative samples, z j represents the feature representation of the j-th sample, sim(.,.) is the similarity function, and τ is the temperature hyperparameter used to adjust the difficulty and sensitivity of contrastive learning.

[0118] The present invention also provides an image enhancement system for low-light environments in complex confined spaces in mines, including a camera, an image preprocessing module, a frequency processing and fusion module, and a postprocessing and optimization module;

[0119] It is installed at the intersection of camera installation roadways and equipment operation points, used to collect original image data in the low-illumination environment of the mine and transmit the image data to the image preprocessing module in real time;

[0120] The described image preprocessing module receives the image data transmitted by the camera, internally integrates an adaptive histogram equalization module and a contrast adjustment module to form a multi-contrast input image, and then outputs it to the frequency processing and fusion module;

[0121] The frequency processing and fusion module uses a convolutional mean filter to separate the high-frequency and low-frequency information of the preprocessed image; the dual-branch processing module strengthens the information according to the frequency characteristics. The high-frequency branch highlights details through three layers of convolution, a channel attention module, and residual connections. The low-frequency branch uses one layer of convolution to focus on the low-frequency content; the frequency fusion module combines through specific convolutional layers and DenseNet layers to fuse and enhance the high-frequency and low-frequency information and outputs it to the post-processing and optimization module;

[0122] The described post-processing and optimization module uses the diffusion model LPDM to remove image noise, optimizes and adjusts through a custom loss function including structural similarity, color constancy, and contrast learning loss, corrects colors, retains structures, and enhances feature representations, and outputs high-quality images that meet the requirements of mine operations.

[0123] During system operation, the camera collects images, the image preprocessing module performs preliminary processing, the frequency processing and fusion module performs in-depth processing, and the post-processing and optimization module improves the output to achieve the enhancement effect of low-illumination images.

[0124] To verify the performance of the present invention, as Figure 7 (a) shows, the low-light image enhancement ability test is carried out using the algorithm of the present invention, and the results are as Figure 7 (b) shows. From the comparison between Figure 7 (a) and (b), it can be seen that the present invention has achieved a significant enhancement effect on low-light images.

Claims

1. An image enhancement method in a complex confined space in a mine with low illumination, characterized in that: The steps include: S1, performing contrast preprocessing on the low-light image through an image preprocessing module; S2, extracting low-frequency information from the preprocessed image through a low-frequency information extraction network, and then subtracting the low-frequency information from the input image to obtain high-frequency information; S3, the low-frequency information and the high-frequency information are fused through the feature fusion network and then input into the image enhancement network for enhancement; S4. Input the enhanced image into the pre-trained LPDM diffusion model to complete post-processing and obtain the final image.

2. The image enhancement method in a complex confined space in a mine under low illumination environment according to claim 1 is characterized in that: The specific process of image preprocessing in S1 is as follows: S1.1, adjusting the low-light image to be processed to a set pixel size; S1.2, divide the low-light image into multiple regions, and perform histogram equalization on each region. The specific steps are as follows: S1.2.

1. Divide the image into K×L sub-blocks B of size m×n ij , where i = 0, 1, ... K-1, j = 0, 1, ..., L-1; S1.2.

2. For each sub-block B ij , calculate its grayscale histogram H ij (k), where k = 0, 1, ... G-1, G is the number of gray levels; S1.2.

3. Calculate each sub-block B ij The cumulative distribution function C ij (k), the formula is: S1.2.

4. For sub-block B ij For each pixel I(x,y), its new gray value in Indicates a round-down operation; S1.

3. After inputting the low-light image i1, adaptive histogram equalization is first performed to obtain i2. Then, based on the image i2, the contrast adjustment coefficient is adjusted within the range of 0.6-0.8 to obtain i3. Then, the contrast of i3 is further adjusted, and the contrast adjustment coefficient is selected within the range of 0.3-0.5 to obtain i4. Finally, i1, i2, i3, and i4 are connected.

3. The image enhancement method in a complex confined space in a mine under low illumination environment according to claim 2 is characterized in that: The specific process of S2 is: S2.

1. Use the convolution mean filter to extract the low-frequency information L(x, y) of the upper image I(x, y), and then use I(x, y) minus the low-frequency information to obtain the high-frequency information; Specifically, let the convolution kernel be K and the size be m×n. The formula for calculating low-frequency information is: The high-frequency information calculation formula is: H(x,y)=I(x,y)-L(x,y); S2.

2. Use a dual-branch frequency enhancement module to enhance the high-frequency and low-frequency information extracted in S2.1, where the formula of the high-frequency branch module is: F high =F+I; F=Conv(Conv(Conv(CA(I)))); The formula of the low-frequency branch module is: F low =DConv(I); Where I is the input feature map, Conv(.) is the convolution operation, CA(.) represents the channel attention module, DConv(.) represents the dilated convolution operation, F represents the features passed through the convolution and attention modules, and F high Represents the enhanced high-frequency information.

4. The image enhancement method in a complex confined space in a mine under low illumination environment according to claim 3 is characterized in that: The specific process of S3 is as follows: splicing the enhanced high-frequency and low-frequency information at the channel, and obtaining the enhanced image through the feature enhancement module, wherein the feature enhancement module is to add a DenseNet between the first network and the second network, and to add a color enhancement attention module; The specific process is: S3.1, Convolutional layer processing: F conv =ReLU(Conv(I,w conv )); S3.2, DenseNet processing: F dense = DenseNet(F conv ) ; S3.3, subsequent convolutional layer and activation function processing: The first convolutional layer: F1 = ReLU (F dense , w1); Second convolutional layer: F2 = ReLU(F1, w2); The third convolutional layer: F3 = Tanh (CEM (F2), w3); Among them, ReLU and Tanh are activation functions, DenseNet(.) is the function of DenseNet, and CEM is the color enhanced attention module.

5. The image enhancement method in a complex confined space in a mine under low illumination environment according to claim 4 is characterized in that: The training of the LPDM diffusion model in S4 is a noise detector, which is divided into two stages: training and reasoning. The specific processes are: S4.1, training phase: sampling noise ∈ from the standard normal distribution, and obtaining the initial real image sample x0 in the mine scene, including mine tunnel image and equipment image, at time step t according to the probability distribution q(x t |x0) Add noise to x0 to obtain the noisy mine image data x t , combined with the mine-related condition information c, the two are added element by element and input into the diffusion model; the diffusion model outputs the predicted noise ∈′, and the loss function L between the real noise ∈ and the predicted noise ∈′ is calculated DM (∈, ∈′) to train the model to improve its accuracy in predicting mine image noise; the loss function L DM The calculation formula is: Among them, θ is the parameter basis of the diffusion model processing input; ∈ θ is the diffusion model function based on parameter θ; α t It is a key variable defined to simplify the derivation and expression of diffusion model related formulas. t =1-β t , where β t ∈(0,1) is the variance parameter that controls the noise added in the forward diffusion process at time step t; S4.2, Reasoning stage: The condition information c related to the mine is first processed by the LLIE backbone network to obtain The result is added element by element to the relevant quantity and then input into the diffusion model; the diffusion model processes the input according to the parameter θ and outputs Then we get the denoised mine image. function based on And parameter s for further processing, where Compare and evaluate the generated results with the real image x0: Among them, α s Variable derived from the variance parameter used to control the noise addition in the forward diffusion process; S4.

3. Using a custom loss function, L ssim is the structural similarity loss, L col is the color constancy loss: L ssim =1-SSIM(x,y); Among them, x, y are the two images to be measured, μ x represents the average value of x pixels, μ y represents the average value of y pixels, σ x and σ y is the variance of the image, σ xy is the covariance of x and y, C1 and C2 are two constants used to maintain stability; Because L ssim Without considering the constraints of color information, the color constancy loss L is used col To correct color distortion: Among them, ρ={(R,G),(R,B),(G,B)}, I p is the average channel value of channel P in the enhanced image; I q is the average channel value of channel q in the reference image; Data collection using contrastive learning: Among them, L i represents the contrastive learning loss of the i-th sample, z i is the feature representation of the current sample, is with z i The corresponding positive sample feature representation, N is the total number of samples including positive samples and negative samples, z j represents the feature representation of the j-th sample, sim(.,.) is the similarity function, and τ is the temperature hyperparameter used to adjust the difficulty and sensitivity of contrastive learning.

6. An image enhancement system in a complex confined space in a mine with low illumination, comprising a camera, characterized in that: It also includes an image pre-processing module, a frequency processing and fusion module, and a post-processing and optimization module; The camera is installed at the intersection of the tunnel and the equipment operation point to collect the original image data in the low-light environment of the mine and transmit the image data to the image preprocessing module in real time; The image preprocessing module receives the image data from the camera, integrates an adaptive histogram equalization module and a contrast adjustment module, forms a multi-contrast input image, and then outputs it to the frequency processing and fusion module; The frequency processing and fusion module uses a convolution mean filter to separate the high-frequency and low-frequency information of the preprocessed image; the dual-branch processing module strengthens the information according to the frequency characteristics, the high-frequency branch is subjected to three layers of convolution and a channel attention module and residual connection to highlight the details, and the low-frequency branch uses a layer of convolution to focus on the low-frequency content; The frequency fusion module combines and enhances high- and low-frequency information through a specific combination of convolutional layers and DenseNet layers, and outputs it to the post-processing and optimization module; The post-processing and optimization module uses the diffusion model LPDM to remove image noise, and optimizes and adjusts the custom loss function including structural similarity, color constancy and contrast learning loss to correct color, preserve structure and improve feature representation, and output high-quality images that meet the needs of mine operations.

Images