Mine image detail enhancement algorithm based on random walk

Abstract

The application belongs to the field of image detail enhancement, and particularly relates to a mine image detail enhancement algorithm based on random walk, which comprises the following steps: a smooth layer I2 and a detail layer I1 of an original image I0 are separated through twice bilinear interpolation, and an initial residual feature Res is calculated; a "starting point" position of random walk, i.e. an original image block, and a "node" position, i.e. a to-be-matched image block, are initialized, and a loss function of the best-matched image block and the original image block is calculated; a better-matched image block is searched and the walking position is continuously updated with the minimum loss function as the target; the globally best-matched image block is gradually converged through continuous iteration of random walk, the update of the residual feature is completed, the updated residual is enlarged and superimposed with the input original image to obtain a final detail-enhanced image. The application innovatively introduces the random walk idea, proposes a new residual feature update mechanism, and optimizes the detail enhancement effect through global search and image block matching.

Description

Technical Field

[0001] This invention belongs to the field of image detail enhancement, specifically relating to an algorithm for enhancing details in underground mine images based on random walk. Background Technology

[0002] When monitoring equipment in mines acquires images, uneven lighting and high levels of airborne impurities often lead to reduced image contrast and significantly decreased detail, hindering the application of intelligent detection technologies. However, by introducing image detail enhancement algorithms, these problems are effectively alleviated. These algorithms not only optimize the overall color distribution of the image but also perform refined processing on local areas, ensuring that key information in the image is clearly presented. The application of this technology greatly improves the usability of underground mine image data, providing a more reliable monitoring method for safe production.

[0003] Most existing algorithms suffer from problems such as gradient inversion artifacts, increased noise, and color deviation. Furthermore, some algorithms exhibit poor generalization ability and cannot adapt to complex scenes. Greedy algorithms, when searching for and matching similar image patches to update residual features, often tend to accept only the best possible state, leading to getting trapped in local optima. This limits the effectiveness of image detail enhancement. These issues collectively restrict the performance of current algorithms in the dark, highly interfering environments of underground mines, affecting the accuracy of safety monitoring equipment in identifying potential hazards and posing a challenge to ensuring safe production in mines. Summary of the Invention

[0004] The technical problem that this invention aims to solve is that existing algorithm models are prone to getting trapped in local optima, which limits the effect of image detail enhancement.

[0005] Therefore, this invention provides an algorithm for enhancing details in underground mine images based on random walks.

[0006] The technical solution adopted by this invention to solve its technical problem is:

[0007] An algorithm for enhancing details in underground mine images based on random walks includes,

[0008] Step 1: Separate the smoothing layer I2 and detail layer I1 of the original image I0 through two bilinear interpolations, and calculate the initial residual features Res;

[0009] Step 2: Initialize the "starting point" position of the random walk (i.e., the original image patch) and the "node" position (i.e., the image patch to be matched), and calculate the loss function between the best matching image patch and the original image patch;

[0010] Step 3: With the goal of minimizing the loss function, find better matching image patches and continuously update the wandering position;

[0011] Step four involves iterating through random walks until the image gradually converges to the globally optimal matching image patch, thus updating the residual features. The updated residual is then amplified and superimposed on the original input image to obtain the final detail-enhanced image.

[0012] Further, in step one, the residual Res of the original image I0 is defined as I0 - I2, and the smoothing layer I2 of the original image I0 is defined as Y × X × I0, where X and Y represent the upsampling matrix and the downsampling matrix, respectively.

[0013] Furthermore, in step two, two positive normalization coefficients α and β are defined in the energy function to force the energy function to focus on the image edge sharpness L. sharpness (s) and texture smoothness L smoothness The energy function is expressed as E(s) = L image (s)+αL sharpness (s)+βL smoothness (s), where L image (s) represents the pixel loss, where s is the vector coordinate of the image patch.

[0014] Furthermore, in step two, the image edge sharpness is represented as: Texture smoothness is represented by L smoothness I(s) = ∑||var(I(s+(Δx,Δy)))-var(I(s))||, where s is the vector coordinate of the image patch, s includes the local features of the local search and the non-local features obtained by the random walk, and I(s) is an image patch centered at s. represents the gradient operator, var represents the image patch variance operator, and Θ is the feasible region for the search and matching process.

[0015] Furthermore, the matched image patch is represented as I(s+(Δx,Δy)), where (Δx,Δy) is a set of coordinate offset vectors.

[0016] Furthermore, in step two, the best matching image patch b * The loss function E(a,b) of the original image patch a * )=|E(a)-E(b * )| is initialized to a very large positive number M.

[0017] Furthermore, in step three, formula x is used. i+1 =x i +λ·u i To simulate the motion of particles, where x' is used. i It is the initial position of the particle, x i+1This represents the particle's position at the next moment. λ controls the step size for each movement. Initially, λ is a large value, and after n consecutive steps without finding the optimal value, its value is updated to λ / 2. i ′ represents the direction of the random walk. First, a two-dimensional direction vector u is randomly generated. i ∈(-1,1) 2 Subsequently, its standardization was achieved. During the iteration process, the search gradually shifts to a more refined approach, eventually converging to the global optimum.

[0018] Furthermore, in step three, during the iteration process, if the loss function value of the image patch at the particle's current position is smaller than the loss value of the best-matching image patch before the walk, then the particle's position x should be recorded at this time. i For x best After swimming, a and b * loss function

[0019] A computer device, comprising:

[0020] processor;

[0021] Memory, used to store executable instructions;

[0022] The processor is configured to read the executable instructions from the memory and execute the executable instructions to implement the random walk-based underground image detail enhancement algorithm as described above.

[0023] A computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to implement the random walk-based image detail enhancement algorithm for underground mining as described above.

[0024] The beneficial effect of this invention is that it innovatively introduces the idea of ​​random walk and proposes a new residual feature update mechanism, which optimizes the detail enhancement effect through global search and image patch matching.

[0025] This invention addresses the problems of poor generalization ability, high complexity, or unsatisfactory results in existing detail enhancement models. It bypasses local extrema in the search and matching process, finding the optimal solution globally to obtain the best-matched image, thereby significantly improving detail enhancement. This results in a marked reduction in artifacts and halos, and demonstrates excellent performance in complex mining environments with uneven lighting and high dust levels, significantly improving the accuracy of safety monitoring equipment. Attached Figure Description

[0026] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0027] Figure 1 This is a flowchart of the random walk-inspired image detail enhancement algorithm for underground mines in this invention.

[0028] Figure 2 This is a block diagram of the algorithm structure of the random walk-inspired image detail enhancement algorithm in mines in this invention.

[0029] Figure 3 This is a simulation diagram of the random walk in this invention.

[0030] Figure 4 This is a schematic diagram comparing the image detail enhancement algorithm of this invention with other existing algorithms to process the image of CUMT-CUMID_train_LR_bicubic dataset 0014 with a four-fold magnification of detail.

[0031] Figure 5 This is a schematic diagram comparing the image detail enhancement algorithm of this invention with other existing algorithms to process the image 0043 of the CUMT-CUMID_train_LR_bicubic dataset with four times the detail magnification.

[0032] Figure 6 This is a schematic diagram comparing the image detail enhancement algorithm of this invention with other existing algorithms to process the image of 0413 in the CUMT-CUMID_train_LR_bicubic dataset with four times the detail magnification. Detailed Implementation

[0033] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0034] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," and "circumferential," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, features defined with "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0035] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0036] Example 1

[0037] A random walk-inspired image detail enhancement algorithm for underground mining utilizes the randomness and global search capabilities of random walks to escape local optima. Specifically, firstly, the original image I0 is subjected to two consecutive bilinear interpolations to obtain I1 and I2, resulting in the residual Res = I0 - I2 of the original image I0. Secondly, the search and matching process of image patches is viewed as minimizing the difference (not limited to pixel differences) in our rewritten energy equation. A random walk algorithm is used for search and matching to obtain the best-matching image patch. Finally, the enhanced residual f is obtained using our novel feature optimization mechanism. new (Res). Finally, the detail-enhanced image I is obtained. enhanced =I0+η×f new (Res). This invention significantly improves image detail enhancement, greatly reduces artifacts and halos, performs excellently in complex mine environments with uneven lighting and high dust levels, and significantly improves the accuracy of safety monitoring equipment.

[0038] A random walk-inspired algorithm for enhancing details in underground mine images includes the following steps:

[0039] The first step involves separating the smoothing layer and detail layer (initial residual features) of the original image patch through two bilinear interpolations.

[0040] Let the original image be I0, and X and Y represent the upsampling and downsampling matrices, respectively, with their scaling ratios being reciprocals of each other. Then, the detail layer I1 = X × I0, and consequently, the smoothing layer I2 of I0 = Y × X × I0. Since bilinear interpolation is inherently coarse, even if the textures of I0 and I2 are very similar, they will still have differences. Therefore, the residual Res of the original image I0 is defined as Res = I0 - I2.

[0041] The second step is to initialize the "starting point" position (original image patch), the "node" position (image patch to be matched), and the loss function for the best matching image patch and the original image patch for the random walk.

[0042] S2.1 Initialization of the random walk starting point. When searching for a match for each image patch, the starting point of the random walk is set to the center point of the current image patch.

[0043] S2.2 The new energy equation is shown in Equation (1). Previous methods for obtaining residuals only utilized features of local regions, and their energy functions only focused on pixel loss L. image (s), as shown in Equation (2). With the introduction of the random walk algorithm, our method can utilize the global features of the image. In fact, for image detail enhancement, in addition to focusing on the pixel loss of the image, its edge sharpness L sharpness (s) and texture smoothness L smoothness (s) also needs attention. In our designed energy function, we define two positive normalization coefficients α and β, forcing the energy function to consider the above-mentioned issues.

[0044] E(s)=L image (s)+αL sharpness (s)+βL smoothness (s) (1)

[0045]

[0046] In the above formula, s is the vector coordinate of the image patch, and I(s) is an image patch centered at s. `var` represents the gradient operator, `var` represents the image patch variance operator, and `Θ` is the feasible region for the search and matching process. It's important to note that `(Δx, Δy)` is a set of coordinate offset vectors, and `I(s + (Δx, Δy))` represents the matched image patch. To leverage the speed and effectiveness of the top-level matching mechanism, and considering the importance of image edges and textures, we define `(Δx, Δy)` as:

[0047]

[0048] It should be noted that the feasible region s in formula (5) includes not only the local features of the local search, but also the non-local features obtained by random walk, which helps us find the global optimal solution.

[0049] S2.3 Based on the newly designed energy function, the best matching image patch b * The loss function E(a,b) of the original image patch a * )=|E(a)-E(b * )| is initialized to a very large positive number M.

[0050] The third step utilizes the characteristics of random walks, randomly selecting the direction of the walk, with the step size controlled by the number of walks. The loss function between the image patch to be matched and the original image patch, and the loss function between the best-matching image patch and the original image patch are calculated. The goal is to minimize the loss function, find a better-matching image patch, and continuously update the walk position.

[0051] S3.1 Escaping Local Optimum. Our method escapes local optima by simulating the disordered motion of particles in space. Previous greedy algorithms are highly likely to stop exploring after finding a local optimum, but the motion of particles in space is unpredictable and will continue to explore after finding a local optimum, thus exploring the entire space.

[0052] In image processing, we use the formula x i+1 =x i +λ·u i To simulate the motion of particles, where x' is used. i It is the initial position of the particle, x i+1 λ represents the particle's position at the next moment, and λ controls the step size of each movement. λ is initially a large value, and after n consecutive attempts without finding the optimal value, its value is updated to λ / 2, gradually transitioning to a finer search, eventually converging to the global optimum. Specifically, u... i ′ represents the direction of the random walk, which is obtained by first randomly generating a two-dimensional direction vector u. i ∈(-1,1) 2 Subsequently, its standardization was achieved.

[0053] S3.2 During the iteration process, the original image patch a and the current best-matching image patch b... * Original image patch a and each image patch b to be searched i The energy loss function is calculated using formulas (6) and (7), respectively.

[0054] E n (s)=E(a,b * )=|E(a)-E(b * (6)

[0055] E n+1 (s)=E(a,b i )=|E(a)-E(b i (7)

[0056] S3.3 Update position. If the loss function value of the image patch at the particle's current position (Equation (7)) is smaller than the loss value of the best-matching image patch before the walk (Equation (6)), then the particle's current position x should be recorded. i For x bestOtherwise, the particle's position should not be recorded. If E n ′(s) represents the values ​​of a and b after the walk. * If the loss function is such that its value can be expressed by formula (8):

[0057]

[0058] The fourth step involves iterating through a certain number of random walks until the image gradually converges to the globally optimal matching image patch, thus updating the residual features. The updated residual is then amplified and superimposed on the original input image to obtain the final detail-enhanced image.

[0059] S4.1 Iterative Update. This process involves repeating step two until the convergence condition is met, i.e., the step size λ is less than the convergence factor ε. λ and ε are parameters that need to be manually adjusted to control the convergence accuracy.

[0060] S4.2 Residual Feature Update. We use our designed random walk-based residual update mechanism f. new (·), using the initial residual features obtained in the first step, calculate the enhanced detail features f. new (Res).

[0061] S4.3 yields the final detail-enhanced image I enhanced =I0+η×f new (Res), where I0 is the original image and η is the detail magnification factor that needs to be manually adjusted to obtain a better visual effect.

[0062] Below is the pseudocode for a detail enhancement algorithm based on random walks:

[0063]

[0064]

[0065] Return: position x best The best matching image patch b at that location

[0066] Note: E(a,b)=|E(a)-E(b)|

[0067] The following section further illustrates the final implementation effect of this invention through specific experiments. Currently, effective image detail enhancement algorithms include GIF, GGIF, WGIF, BFLS, TH, ZF, ILS, and IPRH. The datasets used in the experiments include classic datasets such as Set14, General100, and BSD100, as well as the CUMT-CUMID_train_LR_bicubic dataset. The CUMT-CUMID dataset is a collection of images collected from multiple real mines using professional mining cameras. The image samples in this dataset truthfully and completely record and reflect the impact of different environmental and lighting conditions in the mine on the quality of the images acquired by the equipment. Therefore, the image samples in this dataset are widely applicable to the design, development, and testing of various image processing algorithms for mining environments.

[0068] To quantitatively verify the effectiveness of the proposed algorithm, this embodiment provides a numerical comparison of its objective evaluation metrics with other algorithms. The evaluation metrics we use are the internationally common PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structure Similarity Index Measure). PSNR is an error-sensitive image quality assessment method used to measure the difference between two images. A higher PSNR value indicates a smaller difference between the two images, meaning the corresponding algorithm is superior. SSIM is an image quality assessment metric that measures image similarity from three aspects: brightness, contrast, and structure. The SSIM value ranges from [0,1], and the closer its value is to 1, the smaller the image distortion, and the better the corresponding algorithm.

[0069]

[0070] The calculation methods for PSNR and SSIM are shown in equations (9) and (10); MSE (Mean Square Error) is the mean square error between the original image I and the image K generated by the specific algorithm, and h×w is the resolution of images I and K; μ I and μ K The mean of the image is generated for the original image and a specific algorithm. and The variance σ of the original image and the image generated by a specific algorithm. IK The covariance between the original image and the image generated by the specific algorithm; C1 and C2 are constants set to avoid cases where the denominator is 0.

[0071] Based on the above calculation method, this embodiment provides the quantization comparison results of the proposed algorithm with other mainstream algorithms (including GIF, GGIF, WGIF, BFLS, TH, ZF, and ILS) on the Set14, General100, and BSD100 datasets, as shown in Table 1:

[0072] Table 1 Quantitative Comparison Results

[0073]

[0074] Referring to Table 1, the algorithm proposed in this embodiment performs best in all SSIM tests, indicating that the present invention enhances image details while protecting image structural information most effectively. At the same time, the algorithm proposed in this invention is the best or second best in PSNR tests, indicating that the proposed algorithm can effectively preserve image detail information while reducing noise and distortion, and maintain image clarity and quality well.

[0075] This embodiment also provides the internationally recognized Mean Opinion Score (MOS) to measure the subjective visual performance of each algorithm. The specific process involves selecting a certain number of experts and volunteers without relevant expertise, and randomly selecting the same number of images from each dataset. The images processed by each algorithm are anonymously displayed to these volunteers, and each person assigns a score based on their visual judgment of the image quality. After removing extreme scores, the average score of each image is calculated and the images are sorted from highest to lowest. The MOS scores are shown in Table 2. Our algorithm achieved first or second place in the MOS test, demonstrating the strong robustness of the RW algorithm, its reliable subjective visual performance, and its ability to improve the intuitive visual comfort of most people, effectively enhancing details in various images. It performs particularly well in images from underground mines, making it highly suitable for application in such environments.

[0076] Table 2. MOS score rankings for various algorithms.

[0077]

[0078] To illustrate the effects of the invention through actual visual representation, in Figure 4 , Figure 5 , Figure 6 The algorithm proposed in this embodiment will be compared with the visual performance of some mainstream algorithms on the CUMT-CUMID_train_LR_bicubic dataset. Specifically, Figure 4Image a is from the CUMT-CUMID_train_LR_bicubic dataset 0014. Image a1 is the original image of image a (with four times the detail within the red box). Image a2 is the image processed from image a1 using the BFLS algorithm, with PSNR and SSIM values ​​of 16.89 dB and 0.5475, respectively. Image a3 is the image processed from image a1 using the GIF algorithm, with PSNR and SSIM values ​​of 19.18 dB and 0.6437, respectively. Image a4 is the image processed from image a1 using the TH algorithm, with PSNR and SSIM values ​​of 18.36 dB and 0.6123, respectively. Image 5 shows the image after processing image a1 using the RW algorithm described in this application, with PSNR and SSIM values ​​of 25.12 dB and 0.8692, respectively. Image a6 shows the image after processing image a1 using the GGIF algorithm, with PSNR and SSIM values ​​of 16.82 dB and 0.5777, respectively. Image a7 shows the image after processing image a1 using the ILS algorithm, with PSNR and SSIM values ​​of 18.43 dB and 0.5799, respectively. Image a8 shows the image after processing image a1 using the IPRH algorithm, with PSNR and SSIM values ​​of 23.29 dB and 0.8224, respectively. Figure 5Image b is from the CUMT-CUMID_train_LR_bicubic dataset 0043. Image b1 is the original image with four times the detail (highlighted by the red box) in image b. Image b2 is the image processed from image b1 using the BFLS algorithm, with PSNR and SSIM scores of 15.52 dB and 0.5447, respectively. Image b3 is the image processed from image b1 using the GIF algorithm, with PSNR and SSIM scores of 17.68 dB and 0.6274, respectively. Image b4 is the image processed from image b1 using the TH algorithm, with PSNR and SSIM scores of 18.36 dB and 0.6123, respectively. Image 5 shows the image after processing image b1 using the RW algorithm described in this application, with PSNR and SSIM values ​​of 23.07 dB and 0.8451, respectively. Image 6 shows the image after processing image b1 using the GGIF algorithm, with PSNR and SSIM values ​​of 15.00 dB and 0.5471, respectively. Image 7 shows the image after processing image b1 using the ILS algorithm, with PSNR and SSIM values ​​of 16.52 dB and 0.5852, respectively. Image 8 shows the image after processing image b1 using the IPRH algorithm, with PSNR and SSIM values ​​of 21.17 dB and 0.8011, respectively. Figure 6Image c is from the CUMT-CUMID_train_LR_bicubic dataset 0413. Image c1 is the original image of image c with four times the detail (highlighted by the red box). Image c2 is the image of image c1 processed using the BFLS algorithm, with PSNR and SSIM values ​​of 16.80 dB and 0.5090, respectively. Image c3 is the image of image c1 processed using the GIF algorithm, with PSNR and SSIM values ​​of 19.44 dB and 0.6137, respectively. Image c4 is the image of image c1 processed using the TH algorithm, with PSNR and SSIM values ​​of 19.09 dB and 0.5886, respectively. Image 5 shows the image after processing image c1 using the RW algorithm described in this application, with PSNR and SSIM values ​​of 28.13 dB and 0.8848, respectively. Image c6 shows the image after processing image c1 using the GGIF algorithm, with PSNR and SSIM values ​​of 17.36 dB and 0.5664, respectively. Image c7 shows the image after processing image c1 using the ILS algorithm, with PSNR and SSIM values ​​of 18.48 dB and 0.5617, respectively. Image c8 shows the image after processing image c1 using the IPRH algorithm, with PSNR and SSIM values ​​of 25.77 dB and 0.8391, respectively.

[0079] As can be easily seen from the figure, other algorithms suffer from varying degrees of overfitting, incorrect image color recovery, amplified noise, and halo effects, leading to image distortion. In contrast, the algorithm proposed in this embodiment effectively enhances image details while maintaining a natural overall effect and minimizing distortion. Therefore, the algorithm proposed in this embodiment has significant meaning and practical value in mine image analysis and the modernization of mine construction.

[0080] In summary, the experimental results demonstrate the effectiveness and advantages of this invention. This invention employs a random walk approach to avoid getting trapped in local optima, instead searching for image patches globally, thereby achieving better image detail enhancement.

[0081] Example 2

[0082] This application provides a computer device including a processor and a memory. The memory stores at least one instruction or at least one program, which is loaded and executed by the processor to implement a random walk-based algorithm for enhancing details in underground mine images as provided in the above method embodiments.

[0083] Example 3

[0084] This application embodiment also provides a computer-readable storage medium, which can be disposed in a server to store at least one instruction or at least one program segment for implementing a random walk-based underground mine image detail enhancement algorithm provided in the method embodiment. The at least one instruction or at least one program segment is loaded and executed by the processor to implement the random walk-based underground mine image detail enhancement algorithm provided in the above-described method embodiment. Optionally, in this embodiment, the storage medium can be located in at least one of multiple network servers in a computer network. Optionally, in this embodiment, the storage medium can include, but is not limited to, various media capable of storing program code, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard drive, magnetic disk, or optical disk.

[0085] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined by the scope of the claims.

Claims

1. A method for enhancing details in underground mine images based on random walks, characterized in that, include, Step 1: Separate the smoothing layer I2 and detail layer I1 of the original image I0 through two bilinear interpolations, and calculate the initial residual features Res; Step 2: Initialize the "starting point" position of the random walk (i.e., the original image patch) and the "node" position (i.e., the image patch to be matched), and calculate the loss function between the best matching image patch and the original image patch; Step 3: With the goal of minimizing the loss function, find better matching image patches and continuously update the wandering position; Step 4: Through iterative random walks, gradually converge to the globally optimal matching image patch, complete the update of residual features, amplify the updated residuals and superimpose them with the original input image to obtain the final detail-enhanced image; Among them, the original image patch and the current best matching image patch The loss function is In step two Initialized to a very large positive integer Original image patch and each image patch to be searched The loss function is ; In the energy function, two positive normalization coefficients are defined. a and β This forces the energy function to focus on the sharpness of image edges. and texture smoothness The energy function is expressed as ,in For pixel loss, These are the vector coordinates of the image patch; Image edge sharpness is represented as: Texture smoothness is expressed as ,in, These are the vector coordinates of the image patch. This includes local features obtained from the local search and non-local features obtained from random walks. Therefore A certain image patch centered on The operator representing the gradient is... This represents the operator for calculating the variance of image patches. It is the feasible region for the search and matching process; The matched image patch is represented as ,in, It is a set of coordinate offset vectors. .

2. The method for enhancing details in underground mine images based on random walks according to claim 1, characterized in that, In step one, the original image is defined. residual Original image Smooth layer ,in, and These represent the upsampling matrix and the downsampling matrix, respectively.

3. The method for enhancing details in underground mine images based on random walks according to claim 1, characterized in that, In step three, the formula is used. To simulate the motion of particles, among which It is the initial position of the particle. It is the position of the particle in the next instant. Used to control the step size of each move. Initially, it is a large value, and in succession... n After failing to find the optimal value, its value is updated to , It refers to the direction of the random walk; first, a two-dimensional direction vector is randomly generated. Subsequently, its standardization was achieved. During the iteration process, it gradually transitions to a refined search, eventually converging to the global optimum.

4. The method for enhancing details in underground mine images based on random walks according to claim 3, characterized in that, In step three, during the iteration process, if the loss function value of the image patch at the particle's current position is smaller than the loss value of the best-matching image patch before the walk, then the particle's position should be recorded at this time. for After wandering a and loss function .

5. A computer device, characterized in that, include: processor; Memory, used to store executable instructions; The processor is configured to read the executable instructions from the memory and execute the executable instructions to implement the random walk-based method for enhancing details in underground mine images as described in any one of claims 1 to 4.

6. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, causes the processor to implement the random walk-based method for enhancing details in underground mine images as described in any one of claims 1 to 4.

Images