An image defogging method in a coal mine underground

Abstract

The present application belongs to the technical field of image defogging processing, and discloses a coal mine underground image defogging method, which comprises the following steps: step 1, extracting a brightness component image; step 2, obtaining a gradient image based on the brightness component image, constructing a difference image to identify the area in the original gradient image that is not covered by the inflation operation, then screening out the optimal candidate sub-block and extracting the maximum value of the first 0.1% pixels of the dark channel image of the optimal candidate sub-block as an atmospheric light value; step 3, obtaining a final transmittance image; and step 4, obtaining a dust fog removed image by using an atmospheric scattering model according to the final transmittance image and the atmospheric light value; the present application improves the transmittance estimation accuracy of weak texture and depth of field mutation area, improves the discrimination degree of the weak texture area through a multi-scale weight optimization strategy, makes a better trade-off between detail retention and smoothness of the transmittance, and thus reduces the influence of estimation deviation on the restoration quality.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of image defogging processing, and particularly relates to a coal mine underground image defogging method. BACKGROUND

[0002] The mine monitoring image is degraded under the multiple scattering and absorption of coal dust and water mist. The degradation process is usually expressed by an atmospheric scattering model, that is, the dust-mist image is formed by the superposition of the scene irradiation and the global atmospheric light value, and the transmission item decays exponentially with the depth of field and the scattering coefficient.

[0003] The scarcity of high-quality paired data of mine scenes and the limitation of platform computing power make the physical model method which is interpretable and parameter controllable dominant for a long time. The dark channel prior method is based on the statistical rule of the local minimum channel brightness of natural images, and combines the atmospheric scattering model to estimate the transmittance and atmospheric light of the image. However, due to the uneven illumination and the existence of local strong light source in the underground, the existing method is prone to produce transmittance deviation in the weak texture area covered by dust and mist, and halo and edge diffusion near the strong light, thereby affecting the stability of atmospheric light and detail recovery. SUMMARY

[0004] The purpose of the application is to provide a coal mine underground image defogging method to solve the problem that the existing method is prone to produce transmittance deviation in the weak texture area covered by dust and mist, and halo and edge diffusion near the strong light, thereby affecting the stability of atmospheric light and detail recovery.

[0005] The application adopts the following technical scheme: a coal mine underground image defogging method, comprising:

[0006] Step 1: converting the dust-mist image from the RGB space to the HSV space, and extracting the brightness component graph in the HSV space;

[0007] Step 2: obtaining the atmospheric light value;

[0008] Step 3: obtaining the final transmittance graph;

[0009] Step 4: obtaining the dust-mist removed image by using the atmospheric scattering model according to the final transmittance graph and the atmospheric light value;

[0010] Step 2 specifically comprises:

[0011] Step 201: calculating the gradient graph based on the brightness component graph by the first-order gradient amplitude, and the pixel gray value of the gradient graph represents the local gray change intensity of the image;

[0012] Step 202: performing a gray morphological dilation operation on the gradient graph to enhance the gradient signal continuity of the boundary region;

[0013] Step 203: constructing a difference map based on the gradient map after the expansion operation and the original gradient map, identifying the area in the original gradient map that is not covered by the expansion operation through the difference map, defining a candidate region set in the area not covered by the expansion operation and screening out the optimal candidate sub-block, and then extracting the maximum value of the pixel points ranked in the top 0.1% in the gray value in the dark channel map of the optimal candidate sub-block as the atmospheric light value;

[0014] Step 3 specifically comprises:

[0015] Step 301: based on the dark channel prior theory, obtaining pixel-level dark channel maps, block-level dark channel maps and corresponding pixel-level transmission rate estimation results and block-level transmission rate estimation results under different scales of the dust fog image;

[0016] Step 302: constructing an intermediate response map of different scales based on the difference between the pixel-level dark channel map and the block-level dark channel map;

[0017] Step 303: performing Gaussian blur processing and SMU function enhancement processing on the intermediate response map in sequence to generate an edge-aware weight map;

[0018] Step 304: under a single scale, combining the pixel-level transmission rate estimation result and the block-level transmission rate estimation result to obtain a single-scale fusion result, and weighting and fusing the single-scale fusion results under multiple scales to obtain a preliminary transmission rate map;

[0019] Step 305: taking the brightness component map of the dust fog image as a guide map, introducing a highlight area mask in the guided filtering process, weighting the filtering response of the preliminary transmission rate map, and obtaining a final transmission rate map after refinement processing.

[0020] The beneficial effects of the present application are:

[0021] The present application improves the transmission rate estimation accuracy of weak texture and depth-of-field mutation area, improves the discrimination degree of weak texture area through multi-scale weight optimization strategy, makes a better trade-off between detail retention and smoothness, and thus reduces the influence of estimation deviation on the restoration quality;

[0022] The present application can suppress strong light interference and halo diffusion artifacts, introduce a highlight mask and morphological candidate region extraction in the brightness domain, estimate the atmospheric light value in combination with block search, weaken the interference of point light source on estimation, and reduce edge artifacts;

[0023] The present application can reduce data dependence and computational complexity, without deep learning training under the constraints of lack of paired underground samples and limited equipment computing power, can realize single mine dust fog image defogging and quality enhancement, and can also maintain usable time delay performance and good imaging quality in complex lighting and dust fog environment. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 This is a schematic diagram of the process for obtaining atmospheric light values ​​according to the present invention;

[0025] Figure 2 This is a schematic diagram of the process for obtaining the final transmittance map in this invention;

[0026] Figure 3 This invention illustrates the effect of different scaling factors α on the structural properties of the transmittance map.

[0027] Figure 4 This is a comparison of the restoration results of different algorithms for dust and fog images. Detailed Implementation

[0028] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0029] This invention discloses a method for dehazing images in underground coal mines, which specifically includes four steps.

[0030] Step 1: Convert the dust and fog image from RGB space to HSV space and extract the luminance component map in HSV space.

[0031] Step 2: Obtain atmospheric light values, such as Figure 1 As shown, step 2 specifically includes steps 201-203.

[0032] Step 201: Based on the luminance component map, a gradient map is calculated using the first-order gradient magnitude. The pixel grayscale values ​​of the gradient map represent the intensity of local grayscale changes in the image, specifically:

[0033] ,

[0034] Where (x,y) are pixel coordinates, G x To calculate the gradient along the x-axis from the first-order gradient on the luminance component, G y The gradient of the y-axis is obtained by calculating the first gradient on the luminance component, where V(x,y) is the luminance component of the pixel coordinate point (x,y).

[0035] Since digital images are spatially discretely sampled, the first-order gradient of the brightness component is approximated using a difference method. The central difference method is employed, and its calculation formula is as follows:

[0036] ;

[0037] Among them, G x (x,y) represents the gradient of the (x,y) pixel coordinates along the x-axis, obtained by calculating the first-order gradient on the luminance component. G y (x,y) represents the gradient of the (x,y) pixel coordinates on the y-axis, obtained by calculating the first-order gradient on the luminance component.

[0038] The theoretical basis of the above-mentioned central difference approximation method is the limit definition of the first-order differential, that is:

[0039] ,

[0040] wherein ε is an infinitesimal approaching 0.

[0041] Then, ,

[0042] wherein G(x, y) is a gradient graph, used to describe the intensity of the gray level change of the point.

[0043] Step 202: Perform a gray-scale morphological dilation operation on the gradient graph to enhance the gradient signal continuity of the boundary region and fill the boundary gap.

[0044] In order to exclude the gradient mutation interference of the light source and the strong edge and its neighborhood, perform a gray-scale morphological dilation operation on the gradient graph:

[0045] ,

[0046] wherein B is a structural element, the iteration number of the dilation is set to 1, and G d (x, y) is a gradient graph obtained by performing a gray-scale morphological dilation on the gradient graph G(x, y).

[0047] In order to improve the operation efficiency, the dilation operation is decomposed into one-dimensional sliding maximum filtering in the horizontal and vertical directions. Specifically, each row of the image is first subjected to one-dimensional sliding window filtering, and then each column of the intermediate result is subjected to one-dimensional sliding window filtering with the same parameters.

[0048] The one-dimensional sliding window filtering is realized by optimizing a double-ended queue. Specifically, when traversing the pixels, the gray level value of the current pixel is compared with the pixel gray level value stored at the tail of the double-ended queue. After removing the elements in the queue that are smaller than the gray level value of the current pixel, the index of the current pixel is added to the queue, and the index at the head of the queue that is out of the sliding window range is removed. The index at the head of the queue always corresponds to the maximum gray level value pixel index within the sliding window. While maintaining the real-time update of the extreme value within the sliding window, the calculation complexity of each pixel is reduced to a constant order.

[0049] When the size of the structural element is k x k, the length of the one-dimensional sliding window is also taken as k pixels, so as to keep consistent with the size of the structural element.

[0050] Step 203: constructing a difference map based on the gradient map after the dilation operation and the original gradient map, identifying the region in the original gradient map that is not covered by the dilation operation through the difference map; defining a candidate region set in the region not covered by the dilation operation and screening out an optimal candidate sub-block, and then extracting the maximum value of the pixels ranked in the top 0.1% in the gray value in the dark channel map of the optimal candidate sub-block as the atmospheric light value.

[0051] Firstly, the difference map is constructed to identify the region not covered by the dilation operation, and its expression is:

[0052]

[0053] Wherein, D(x, y) is the difference map.

[0054] The candidate region set is defined as:

[0055]

[0056] Wherein, the threshold K is 2 / 255, and C is the candidate region set.

[0057] In step 203, when the optimal candidate sub-block is screened out, the candidate region is first divided into four sub-blocks layer by layer, and then a scoring function is constructed based on the brightness consistency, and the candidate sub-block with the highest scoring result is taken as the optimal candidate sub-block. The scoring function S is specifically:

[0058]

[0059] Wherein, μ is the average value of the candidate sub-block in the brightness channel, and σ is the brightness standard deviation. The higher the score is, the higher and more uniform the brightness in the region is, and the more likely it is to be the atmospheric light estimation region.

[0060] In the optimal candidate sub-block, the top 0.1% of the pixel points in the dark channel map are taken as the atmospheric light estimation region, and the maximum value in the region is output as the atmospheric light value A.

[0061] Step 3: obtaining the final transmittance map, as shown in Figure 2 Step 3 specifically includes steps 301-305.

[0062] Step 301: based on the dark channel prior theory, obtaining the pixel-level dark channel map, the block-level dark channel map under different scales of the dust image, and the pixel-level transmittance estimation result and the block-level transmittance estimation result corresponding to the scale.

[0063] Step 302: constructing an intermediate response map of different scales based on the difference between the pixel-level dark channel map and the block-level dark channel map. The calculation formula of the intermediate response map is specifically:

[0064] ​​​​

[0065] Where, d pi This is a pixel-level dark channel image, d pa Block-level dark channel diagram; D t This is an intermediate response diagram.

[0066] Introducing a highlight area mask M light And by using the brightness component of the dust and fog image to set a high brightness threshold, potential overexposed areas can be identified:

[0067] ,

[0068] Among them, M light θ is the mask for the highlighted area, θ is the brightness threshold, and V is the brightness component value, used to distinguish between normal and overexposed areas. By statistically analyzing the histogram distribution of dust and fog images in the brightness channel, the top 0.1% of brightness values ​​are selected as the brightness threshold θ.

[0069] Step 303: Perform Gaussian blurring and SMU function enhancement on the intermediate response map in sequence to generate an edge-aware weight map.

[0070] Gaussian blur and SMU function are used to smooth and enhance the intermediate response map to generate an edge-aware weight map, which can be specifically represented as follows:

[0071] ,

[0072] Among them, D f ω is the edge-aware weight map, ω is the smoothing control parameter with a value of 1; α is the scaling factor, and GaussianBlur is the Gaussian blur.

[0073] The SMU function stands for Smooth Maximum Unit function.

[0074] like Figure 3 As shown, the standard deviation of transmittance and the number of visible edges exhibit a non-monotonic trend with the increase of the scaling factor. The detail representation and structural contrast of image edges reach their optimal values ​​within the range α ∈ [0.5, 0.6]. When α further increases, the structural enhancement effect of the transmittance map tends to saturate, even leading to weakening of local details and a reduction in overall continuity. Therefore, the preferred value for α is 0.6.

[0075] This invention introduces an edge-aware weighted graph D f This is used to measure the difference in transmittance estimation results at different scales. Traditional tanh nonlinear functions have limited response capabilities in edge regions, and are prone to saturation, especially when processing local brightness abrupt changes or structural variations in images, affecting the accuracy of the weight map. Therefore, this invention introduces the SMU function, which enhances edge perception capabilities through a smoothing maximization mechanism.

[0076] Step 304: At a single scale, combine the edge-aware weight map to fuse the pixel-level transmittance estimation result and the block-level transmittance estimation result to obtain a single-scale fusion result, and weight fuse the single-scale fusion results at multiple scales to obtain a preliminary transmittance map.

[0077] First, fuse the pixel-level and block-level transmittance estimation results at a single scale as follows:

[0078]

[0079] wherein, is a regularization coefficient for avoiding division by zero; t i mof is a single-scale fusion result; t pi is a pixel-level transmittance estimation result, t r pa is a block-level transmittance estimation result at the r-th scale.

[0080] In order to fully utilize the advantages of estimation at different scales and weaken the problem that small scales are easily disturbed by noise in thick fog areas, the single-scale fusion results at each scale are weight fused. Considering that large-scale estimation has better robustness in heavy fog areas, the present application adopts a scale weight based on exponential decay, specifically as follows:

[0081]

[0082] wherein, W i is the weight of the i-th scale; r i is the patch size corresponding to the i-th scale, r j is the patch size corresponding to the j-th scale, i and j are both scale indexes, r N is the maximum patch size, is a weight decay factor, which is set to 0.138 according to experience; and N is the total number of scales. The larger the patch size, the higher the corresponding weight.

[0083] Therefore, the preliminary transmittance map is obtained by weight fusing the single-scale fusion results at multiple scales as follows:

[0084]

[0085] wherein, t fused is the preliminary transmittance map.

[0086] Step 305: Take the luminance component map of the dust fog image as a guide map, fuse a highlight area mask in the guide filtering process, weight the filtering response of the preliminary transmittance map, and obtain a final transmittance map t s (x) after refinement.

[0087] ​​​Step 4: Obtain the haze-free image by using the atmospheric scattering model according to the final transmittance map and the atmospheric light value, specifically: combine the obtained atmospheric light value A, the final transmittance map t s (x), and perform haze removal by atmospheric scattering model inversion:

[0088] ,

[0089] wherein J(x) is scene radiance, also the haze-free image; A is the atmospheric light value; t0 is a lower limit value set for transmittance; t s (x) is the final transmittance map, and I(x) is the haze image. Since t s (x) is too small, J(x) will be too large, and the overall haze-free image will transition to the white field, therefore max[t s (x), t0] indicates setting a lower limit value for transmittance, and the value is 0.1.

[0090] Compared with the method of using convolutional neural network (CNN) or generative adversarial network (GAN) for mine image haze removal in the prior art, such deep learning methods usually require a large amount of paired high-quality training data to learn image features in complex environments. However, in the mine scene, high-quality data is difficult to obtain, and there is a significant domain offset between synthetic haze images and real images, thereby affecting the generalization ability of the model. In addition, the deep learning method has high computational complexity and is not easy to deploy on devices with limited computing power. In contrast, the present application can obtain relatively stable haze removal effect under various lighting conditions without relying on training data, and has good recovery performance in non-uniform illumination and complex detail areas.

[0091] Compared with the image enhancement algorithm and image filtering technology in the prior art, although these methods in the prior art have low computational complexity, are suitable for real-time image processing, and do not rely on large-scale data. However, since these algorithms mainly rely on heuristic rules, they often cannot fully consider the local illumination differences and transmittance changes in the image, so the haze removal effect in complex scenes is limited, especially in strong light sources and weak texture areas.

[0092] Compared with the wavelet transform method in the prior art, wavelet transform decomposes the image into detail and approximation images of different scales, restores the transmittance in the low-frequency part, and restores the details in the high-frequency part. Although wavelet transform can effectively separate the low-frequency and high-frequency components in the image, thereby realizing haze removal and detail enhancement. However, the wavelet transform has limited ability to handle local details in extreme strong light and complex environments, and cannot handle local details as well as the combination of multi-scale and edge perception in the present application.

[0093] Embodiment: Comparison of the restoration results of the haze image by using different algorithms, and the results are as follows: Figure 4and Table 1, Figure 4 The scene 1 and the scene 2 are both coal wall cutting areas near the drum of the coal mining machine in the fully mechanized coal mining face, the scene 3 is a support channel area where the support in the middle of the fully mechanized coal mining face and the conveyor in the working face are located, and the scene 4 is a transfer point of the transportation roadway and a concentrated area of equipment.

[0094] From Figure 4 It can be seen that:

[0095] The DCP algorithm has good defogging effect, but is darker than the dust fog image, and has insufficient brightness and detail recovery in the coal wall and coal falling area in the scene 1.

[0096] The IDE algorithm has limited defogging effect on the thick fog area, but has improved detail and brightness recovery in the small amount of fog or no fog area in the scene 3; but the overall brightness of the restored image is too high due to the interference of the point light source and the thick fog, and the halo effect is significant, resulting in poor visual effect.

[0097] The CEEF algorithm effectively enhances the contrast of the image, especially in the close-up area of the scene 1 and the scene 4, but the dust fog in the long-range part is still obvious, the defogging effect is relatively balanced, and part of the details are recovered.

[0098] The ROP algorithm has similar performance to the IDE algorithm, but performs better in retaining the details and textures of the cutting head area and the coal wall in the scene 2.

[0099] The brightness of the SLP algorithm is basically the same as that of the original image, but the defogging effect in the thick fog area is general, and the defogging in the thin fog area is not complete.

[0100] After the defogging processing by the algorithm of the embodiment, the brightness and details of the coal wall and the remote working face in the scene 1 are significantly improved, and the profiles of the coal mining machine and the conveying equipment, which are originally blocked by the dust fog, are clearer; although the cutting head rotates at high speed and the illumination is complex in the scene 2, the texture structure of the cutting head and the coal wall can still be effectively maintained after enhancement, the brightness and dark transition is more natural, and the traditional algorithm is avoided to appear overexposure and block noise, so that the surface details of the working face equipment are clearer; the details and textures of the coal wall are also effectively recovered, and the brightness and dark transition in the local area is more uniform, which is conducive to observing the working face environment; the strong light in the top area in the scene 3 is obviously suppressed, and the structures such as the support and the conveyor can still be clearly identified under weak light conditions, so as to avoid glare caused by excessive divergence of light source, it can be seen that the embodiment can significantly suppress the intensity of the light source, so that the brightness around the top light does not excessively diverge, thereby effectively retaining the detail information of the support, the conveyor and other structures; the contrast of the small-scale structures such as the equipment and the cable at the end of the roadway in the scene 4 is further improved, and the overall recognition degree is obviously better than that of other algorithms.

[0101] In summary, the image processed by the algorithm of the embodiment has more natural and ideal effects in detail recovery, brightness balance and structural integrity.

[0102] Table 1. Image quality evaluation results after different algorithms are used for defogging

[0103]

[0104] In Table 1, FADE is a haze concentration evaluation index, σ is a standard deviation, H is an information entropy, AG is an average gradient, NIQE is a natural image quality evaluation index, and BRISQUE is a no-reference spatial domain image quality evaluation index.

[0105] Among them, FADE, NIQE and BRISQUE are all no-reference image quality evaluation indexes, and the smaller the value is, the less the image haze is, the lower the distortion is, and the better the subjective quality is.

[0106] Among them, σ, H and AG respectively represent image contrast, information amount and definition, and on the premise of no oversaturation, the larger the value is, the better the image quality is.

[0107] In Table 1, the algorithm DCP is derived from Kaiming He, Jian Sun and Xiaoou Tang, Single image haze removal using dark channel prior, IEEE Conference on Computer Vision and Pattern Recognition, Miami, FL, USA, 2009, pp. 1956-1963, doi: 10.1109 / CVPR.2009.5206515.

[0108] In Table 1, the algorithm IDE is derived from M. Ju, C. Ding, W. Ren, Y. Yang, D. Zhang and Y. J. Guo, IDE: Image Dehazing and Exposure Using an Enhanced Atmospheric Scattering Model[J], IEEE Transactions on Image Processing, vol. 30, pp. 2180-2192, 2021.

[0109] In Table 1, the algorithm CEEF is derived from X. Liu, H. Li and C. Zhu, Joint Contrast Enhancement and Exposure Fusion for Real-World Image Dehazing[J], IEEE Transactions on Multimedia, vol. 24, pp. 3934-3946, 2022.

[0110] In Table 1, the algorithm ROP is derived from J. Liu, R. W. Liu, J. Sun and T. Zeng, Rank-One Prior: Real-Time Scene Recovery[J], IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 7, pp. 8845-8860, 1 July 2023, doi: 10.1109 / TPAMI.2022.3226276.

[0111] In Table 1, the algorithm SLP is derived from P. Ling, H. Chen, X. Tan, Y. Jin and E. Chen, Single Image Dehazing Using Saturation Line Prior[J], IEEE Transactions on Image Processing, vol. 32, pp. 3238-3253, 2023.

[0112] As can be seen from Table 1, the embodiment has the best overall performance in the FADE index, better than the ROP and SLP methods, indicating that the embodiment has more advantages in removing image fog and improving overall clarity. Although the DCP algorithm can still obtain a lower FADE value in some scenes, the corresponding information entropy and average gradient values are generally low, indicating that it has obvious deficiencies in image detail preservation and limited adaptability to complex low-illumination environments.

[0113] In contrast, the IDE and ROP methods can improve local contrast, but have shortcomings in color restoration and naturalness preservation. The CEEF algorithm has strong detail recovery capability, but its FADE and standard deviation indicators are lower than those of the embodiment, indicating that it has problems such as poor clarity or local over-enhancement. The SLP algorithm performs well in structural contrast enhancement, but the overall image brightness is insufficient, affecting the final visual effect.

[0114] In the NIQE and BRISQUE indexes for measuring the naturalness of the image, the embodiment also performs excellently, can effectively suppress the fog interference while maintaining the natural visual effect of the image, and is suitable for image defogging tasks in low-contrast and complex lighting environments.

[0115] The complexity and implementation efficiency of the defogging calculation directly determine the response time of the subsequent detection. To verify whether the embodiment has good defogging efficiency, the average running time of the DCP defogging algorithm, the CEEF defogging algorithm, the ROP defogging algorithm, the IDE defogging algorithm and the embodiment on a single image is counted. To quantify the efficiency, the average running time of DCP, ROP, CEEF, IDE and the embodiment is counted under the same platform and the same input resolution (400x600).

[0116] Table 2 Average running time of different algorithms on a single image

[0117]

[0118] The algorithms DCP, CEEF, ROP and IDE in Table 2 and the corresponding symbols in Table 1 have the same source.

[0119] As can be seen from Table 2, the embodiment only needs 0.023 seconds under the resolution of 400x600, and the running time is significantly reduced. This performance improvement is due to the following reasons:

[0120] (1) The gray-scale morphological dilation and dark channel are both implemented by using a one-dimensional sliding window-based local convolution kernel. The originally complex operation that grows with the window radius is optimized to be proportional to the number of image pixels, thereby making the overall calculation not sensitive to the window size and suitable for any resolution, thus improving the speed.

[0121] (2) The multi-scale fusion strategy with exponential decay weighting is introduced to effectively reduce the redundancy of large-scale patches.

[0122] (3) It does not rely on deep learning models or GPU acceleration, but can be completed based on traditional image processing operators, and can be stably run on a CPU platform, has good embedded adaptation ability, and is suitable for resource-constrained mine monitoring systems.

[0123] The above are only preferred embodiments of the present application and are not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for dehazing images in underground coal mines, characterized in that, include: Step 1: Convert the dust and fog image from RGB space to HSV space, and extract the luminance component map in the HSV space; Step 2: Obtain atmospheric light values; Step 3: Obtain the final transmittance map; Step 4: Obtain the dust-free image using an atmospheric scattering model based on the final transmittance map and atmospheric light values; Step 2 specifically includes: Step 201: Based on the brightness component map, a gradient map is calculated using the first-order gradient magnitude. The pixel grayscale values ​​of the gradient map represent the intensity of local grayscale changes in the image. Step 202: Perform a grayscale morphological dilation operation on the gradient map to enhance the continuity of the gradient signal in the boundary region; Step 203: Construct a difference map based on the gradient map after the dilation operation and the original gradient map, and identify the regions in the original gradient map that are not covered by the dilation operation through the difference map; define a set of candidate regions in the regions not covered by the dilation operation and select the optimal candidate sub-block, and then extract the maximum value of the top 0.1% of the pixels in the dark channel image of the optimal candidate sub-block as the atmospheric light value; Step 3 specifically includes: Step 301: Based on the dark channel prior theory, obtain pixel-level dark channel images and block-level dark channel images of the dust and fog image at different scales, as well as pixel-level and block-level transmittance estimation results at the corresponding scales. Step 302: Construct intermediate response maps of different scales based on the differences between the pixel-level dark channel map and the block-level dark channel map; Step 303: Perform Gaussian blurring and SMU function enhancement processing sequentially on the intermediate response map to generate an edge-aware weight map; Step 304: At a single scale, combine the edge-aware weight map to fuse the pixel-level transmittance estimation results and the block-level transmittance estimation results to obtain a single-scale fusion result, and perform weighted fusion of the single-scale fusion results at multiple scales to obtain a preliminary transmittance map. Step 305: Using the brightness component map of the dust and fog image as a guide map, a high-brightness region mask is incorporated into the guided filtering process to weight the filtering response of the preliminary transmittance map. After refinement, the final transmittance map is obtained.

2. The method for dehazing images in underground coal mines according to claim 1, characterized in that, In step 203, when selecting the optimal candidate sub-block, the candidate region is first divided into four sub-blocks layer by layer. Then, a scoring function is constructed based on brightness consistency, and the candidate sub-block with the highest score is selected as the optimal candidate sub-block. The calculation formula for the scoring function S is as follows: ， Where μ is the average value of the candidate sub-blocks in the luminance channel, and σ is the luminance standard deviation.

3. The method for dehazing images in underground coal mines according to claim 1, characterized in that, The calculation formula for the edge-aware weight map in step 303 is as follows: ， Among them, D f This is the edge-aware weight map, where ω is the smoothing control parameter; α is the scaling factor; GaussianBlur is the Gaussian blur; D t This is an intermediate response diagram; ， Where, d pi This is a pixel-level dark channel image, d pa Block-level dark channel diagram; ， Among them, M light θ is the mask for the highlighted area, θ is the luminance threshold, and V is the luminance component value.

4. The method for dehazing images in underground coal mines according to claim 1, characterized in that, In step 304, when performing weighted fusion of single-scale fusion results at multiple scales, the fusion is based on exponentially decaying scale weights. The formula for calculating the scale weights is as follows: ， Among them, W i r represents the weight at the i-th scale; i Let r be the patch size corresponding to the i-th scale. j Let r be the patch size corresponding to the j-th scale, where i and j are scale indices. N Maximum patch size, is the weight decay factor, and N is the total number of scales.

5. The method for dehazing images in underground coal mines according to claim 1, characterized in that, The atmospheric scattering model is as follows: ， Where J(x) is the image after dust and fog removal; A is the atmospheric light value; t0 is the lower limit of transmittance; t s I(x) is the final transmittance map, and I(x) is the dust and fog image.

6. The method for dehazing images in underground coal mines according to claim 1, characterized in that, The grayscale morphological dilation operation in step 202 is decomposed into one-dimensional sliding maximum filtering in the horizontal and vertical directions. Specifically, one-dimensional sliding window filtering is first performed on each row of the gradient map, and then one-dimensional sliding window filtering with the same parameters is performed on each column of the intermediate result after row filtering.

7. The method for dehazing images in underground coal mines according to claim 1, characterized in that, The grayscale morphological dilation operation in step 202 is iterated once.

Images