Image detail enhancement algorithm based on PIMO projection iterative optimization

Abstract

The invention discloses an image detail enhancement algorithm based on PIMO projection iterative optimization. The image detail enhancement algorithm comprises the following steps: S1, image preprocessing and multi-scale decomposition; s2, multi-dimensional feature calculation is carried out, an intensity-gradient-texture triple feature space is constructed, and a matching evaluation basis is provided; s3, performing PIMO projection iterative optimization search, searching an optimal detail block through an operator, and reconstructing an enhanced detail layer; s4, performing detail fusion and enhanced superposition, and performing adaptive superposition on the reconstructed detail layer and the original image to generate an enhanced image; and S5, quality evaluation and result output: quantitatively evaluating an enhancement effect and exporting a processing result. The mine image enhancement method can break through the limitation of a traditional method in the aspects of detail-noise balance, calculation efficiency, self-adaptive capability and the like, provides an efficient, accurate and self-adaptive technical path for mine image enhancement, and realizes accurate search and efficient enhancement of detail information.

Description

Technical Field

[0001] This invention belongs to the field of image detail enhancement technology, specifically relating to an image detail enhancement algorithm based on PIMO projection iterative optimization. Background Technology

[0002] With the widespread application of digital imaging technology in fields such as industrial inspection, security monitoring, medical imaging, and remote sensing, image quality directly affects the accuracy of subsequent analysis and decision-making. In the construction of smart mines, image processing technology is crucial for applications such as mine monitoring, equipment fault diagnosis, and environmental assessment. However, the mine environment is typically characterized by low illumination, high noise, and complex interference, leading to loss of detail, blurred edges, and reduced contrast in acquired images, severely restricting the effective extraction and utilization of image information.

[0003] Traditional image detail enhancement methods, such as histogram equalization, grayscale transformation, and filtering denoising, while effective in specific scenarios, generally suffer from poor adaptability, over-enhancement, or the introduction of artifacts. Especially in complex environments with low light and high noise, such as mines, underground spaces, and nighttime monitoring, these methods struggle to effectively suppress noise while enhancing details, often leading to further image quality degradation and limiting their practical application in smart mines. For example, in mine equipment monitoring, over-enhancement may amplify sensor noise and mask key fault characteristics; in environmental assessment, artifact interference may lead to misjudgments of terrain details, affecting safety decisions.

[0004] In recent years, deep learning-based image enhancement methods have gradually attracted attention, achieving significant results in areas such as image super-resolution and denoising through models like convolutional neural networks (CNNs). However, the application of these methods in smart mines faces two major challenges: first, data annotation is difficult, as it is hard to acquire large amounts of image data in the special scenarios of mines, and the annotation process is time-consuming and labor-intensive; second, computational resources are high, as deep learning models typically require high-performance GPUs, resulting in high training and inference complexity, making them difficult to deploy on edge devices or real-time monitoring systems in mines. Furthermore, deep learning methods are highly dependent on multiple sample data, exhibiting limited generalization ability in low-light, high-noise mine scenarios where samples are scarce, and are prone to detail distortion or structural deformation.

[0005] Therefore, the image detail enhancement technologies currently used in smart mines still have the following prominent problems: 1) The detail-noise balance problem: Traditional methods and some optimization algorithms have difficulty effectively distinguishing between real structure and noise signals when enhancing details, often resulting in noise amplification or detail smoothing, which affects the reliability of subsequent analysis.

[0006] 2) Computational efficiency and real-time bottlenecks: Complex optimization algorithms or deep learning models have high computational overhead and take a long time to process a single high-resolution image, which cannot meet the real-time processing requirements of mine monitoring systems.

[0007] 3) Insufficient adaptability: Existing methods mostly rely on fixed parameters or pre-trained models, which are difficult to adapt to the image quality fluctuations caused by multiple factors such as changes in lighting, dust interference, and equipment vibration in the mining environment.

[0008] 4) Low efficiency of high-dimensional feature search: Optimization methods based on global search have high computational complexity and slow iterative convergence in the high-dimensional feature space of images, making it difficult to achieve efficient detail matching with limited resources. Summary of the Invention

[0009] The purpose of this invention is to provide an image detail enhancement algorithm based on PIMO projection iterative optimization, which can overcome the limitations of traditional methods in terms of detail-noise balance, computational efficiency and adaptability, and provide an efficient, accurate and adaptive technical path for mine image enhancement, realizing accurate search and efficient enhancement of detail information.

[0010] To achieve the above objectives, this invention provides an image detail enhancement algorithm based on PIMO projection iterative optimization, comprising the following steps: S1. Image preprocessing and multi-scale decomposition; S2. Multi-dimensional feature calculation constructs a triple feature space of "intensity-gradient-texture" to provide a basis for matching evaluation; S3 and PIMO projection iterative optimization search find the optimal detail block through operators and reconstruct the enhanced detail layer; S4. Detail Blending and Enhancement Overlay: The reconstructed detail layer is adapted and overlaid with the original image to generate an enhanced image; S5. Quality assessment and result output: Quantitative evaluation enhances the effect and exports the processing results.

[0011] As a further aspect of the present invention: S1 image preprocessing and multi-scale decomposition specifically includes: S1.1 Perform format standardization processing on the input image; S1.2. Multi-scale scaling is performed through bilinear interpolation to generate a magnified image L1 and a smoothing layer L2 restored to the original size. The difference between the original image and the smoothing layer L2 is calculated to obtain the initial detail layer H0, thus achieving smoothing-detail separation of the image. S1.3. Use the copy-fill method to expand the boundaries of images at each scale to avoid artifacts in the edge region during feature calculation and block matching.

[0012] As a further aspect of the present invention: S2 multi-dimensional feature calculation specifically includes: S2.1 Calculate the horizontal and vertical gradients of smoothing layers at each scale based on the Sobel operator, and integrate the bidirectional gradient information through the gradient magnitude formula to construct gradient features; S2.2. Perform convolution operation on the grayscale image using a Laplacian convolution kernel to generate a texture response map and extract texture features; S2.3 Combine the original intensity information of the image with gradient and texture features to construct a triple feature space of "intensity-gradient-texture", which provides a comprehensive basis for the similarity evaluation of subsequent matching blocks.

[0013] As a further aspect of the present invention: S3 PIMO projection iterative optimization search, specifically including: S3.1 Initialize the population and parameters: Randomly generate an initial population of size N, with the image space coordinates corresponding to the solution vectors, and the search dimension di; set the maximum number of iterations. Max iter Define the Jacobian matrix J Levi's flight parameters β Initialize elite solution X best Used for iterative guidance; S3.2, Block Sampling and Matching Criterion Definition: The magnified image L1 is sampled in 5×5 blocks with a step size of 2 pixels. The intensity, gradient and texture triple features of each sampling block are extracted as reference blocks. The fitness function is defined as the sum of the absolute errors of the triple features of the reference block and the candidate block, which quantifies the matching similarity. S3.3, Operator iteration execution, including: ① Residual-guided projection: Reference individuals are adaptively selected based on the residual size, and position updates are achieved through stochastic gradient calculation and Jacobian projection; ②Double random projection: Randomly select two individuals in the population as reference points, and generate update directions through two gradient combination methods to enhance global exploration capabilities; ③ Weighted random projection update: Two update paths are designed, and the random weights and iteration progress are adaptively selected to balance the breadth of exploration and the accuracy of development; ④ Levy Flight Guidance Projection: Dynamically Adjusting Trigger Probability Based on Hyperbolic Tangent Function O By using Levy's flight to generate long-distance jump steps, the algorithm can be helped to find local optima. S3.4 Boundary Processing and Elite Preservation: A reflection boundary strategy is used to limit the solution space range, ensuring that the updated position is within the effective image area. After each iteration, the individual with the best fitness is selected as the elite solution. X best This information is then passed on to the next iteration to ensure the correctness of the optimization direction. S3.5, Detail layer reconstruction: The detail information corresponding to the best matching block obtained by all iterative search is accumulated and averaged, and the region with a count of zero is assigned a value of 1 to obtain the enhanced detail layer Residual.

[0014] As a further aspect of the present invention: the reflection boundary strategy is as follows: If the updated position is lower than the lower bound, the reflection position is calculated using "2 × lower bound - updated position"; If the updated position is higher than the upper bound, the reflection position is calculated using "2 × upper bound - updated position"; If it is within the valid range, then maintain the original position; The lower and upper bounds represent the minimum and maximum valid values ​​of the image coordinates, respectively.

[0015] As a further aspect of the present invention: S4 detail blending and enhancement overlay specifically includes: S4.1 Align the reconstructed detail layer Residual with bilinear interpolation according to the original image size to ensure spatial matching between detail information and the original image; S4.2 Perform enhancement overlay operations on the three RGB channels separately; S4.3. The superimposed pixel values ​​are truncated and limited to the effective pixel range of 0-255, and finally converted into an 8-bit unsigned integer format to obtain the enhanced image.

[0016] As a further aspect of the present invention: S5 quality assessment and result output specifically includes: S5.1 Calculate the quantitative evaluation index using the Y channel of the YCbCr color space; S5.2 supports manual saving of single images and automatic export of batch datasets. During batch processing, an "Enhanced Images" subfolder is generated to store all enhanced image files. At the same time, index statistics files in text and data formats are generated to record the PSNR, SSIM values ​​and overall average index of each image.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Excellent search efficiency: The four core operators of the PIMO algorithm work together, combined with dynamic decay step size and Levy flight guidance, to achieve efficient exploration and accurate development in the solution space. Compared with traditional optimization algorithms, the search speed is improved by more than 30%, and the processing time of a single image is ≤0.8 seconds, which meets the requirements of real-time processing.

[0018] 2. Precise Detail Enhancement: Based on triple feature matching of "intensity-gradient-texture" and PIMO optimized search, it can effectively distinguish between real details and noise. The SSIM value of the enhanced image is ≥0.96, avoiding the artifacts and noise amplification problems caused by over-enhancement in traditional methods.

[0019] 3. Strong Adaptability: Through adaptive trigger probability, dynamic step size, random weight and other parameter adjustment mechanisms, it can adapt to different types and qualities of images without manual intervention, and maintain stable performance in multiple scenarios such as smart mines and equipment monitoring.

[0020] 4. Good adaptability to high dimensions: The algorithm's time complexity is (where T is the number of iterations, N is the population size, and D is the search dimension), and its growth trend with the increase of image resolution is gradual. It still maintains high efficiency in high-resolution image processing and reduces the consumption of computing resources compared with deep learning methods.

[0021] 5. Wide range of applications: It supports real-time processing of single images and automated enhancement of batch datasets. The output includes enhanced images and quantitative indicators, which can directly provide high-quality data support for subsequent tasks such as image analysis, target detection, and fault diagnosis. It is suitable for multiple fields such as smart mines, equipment monitoring, and environmental monitoring.

[0022] This invention innovatively introduces the PIMO projection iterative optimization algorithm into the field of image detail enhancement, constructing a collaborative framework of multi-scale features and metaheuristic optimization, overcoming the limitations of traditional enhancement methods in detail-noise balance and adaptability. This algorithm requires no large-scale labeled data or complex model training, boasts high computational efficiency, and produces natural enhancement results, providing a new technical path for image detail enhancement. It has significant promotional value and broad application prospects in practical applications such as intelligent mine construction and industrial inspection. Attached Figure Description

[0023] Figure 1 This is a flowchart of the image detail enhancement algorithm based on PIMO projection iterative optimization according to the present invention.

[0024] Figure 2 This is a block diagram of the image detail enhancement algorithm based on PIMO projection iterative optimization according to the present invention.

[0025] Figure 3 This is a comparison diagram of the present invention with other algorithms.

[0026] Figure 4 This is the first result of a quantitative comparison between this invention and other algorithms.

[0027] Figure 5 This is the second result of a quantitative comparison between this invention and other algorithms. Detailed Implementation

[0028] The invention will now be further described with reference to the accompanying drawings.

[0029] like Figure 1 As shown, an image detail enhancement algorithm based on PIMO projection iterative optimization includes the following steps: S1. Image preprocessing and multi-scale decomposition: The input image is format-converted and type-unified, color images are converted to grayscale images, and then converted to double type to avoid truncation of negative values ​​during the calculation process.

[0030] Specifically, this includes: S1.1 Standardize the format of the input image. Color images retain the RGB three-channel structure, and grayscale images are converted to a three-channel format to unify the processing flow and ensure compatibility of subsequent operations. S1.2. Perform multi-scale scaling through bilinear interpolation to generate a magnified image L1 (preferably a magnified image at 1.25 times the scale) and a smoothing layer L2 restored to the original size. Calculate the difference between the original image and the smoothing layer L2 to obtain the initial detail layer H0, thereby achieving smoothing-detail separation of the image. S1.3. Use the copy-fill method to expand the boundaries of images at each scale. The fill size is the size of the search window. It is preferred to set the fill window size to 20 pixels to avoid artifacts in the edge area during feature calculation and block matching.

[0031] S2. Multi-dimensional feature calculation: Through image scaling and interpolation, the preprocessed image is decomposed into a basic layer and a high-frequency detail layer, constructing a triple feature space of "intensity-gradient-texture" to provide a basis for matching evaluation.

[0032] Specifically, it includes: S2 multi-dimensional feature calculation specifically includes: S2.1 Calculate the horizontal and vertical gradients of smoothing layers at each scale based on the Sobel operator, and integrate the bidirectional gradient information through the gradient magnitude formula to construct gradient features; the gradient magnitude formula is as follows: ; in: : Gradient magnitude of a pixel; The horizontal gradient value is calculated by convolving the image with the Sobel operator in the horizontal direction. The vertical gradient value is calculated by convolving the image with the Sobel operator in the vertical direction. S2.2. Perform convolution operations on the grayscale image using a Laplacian convolution kernel to generate a texture response map and extract texture features; the formula is as follows: ; in: Input image Texture feature map; : Laplace operator kernel, specifically ; Grayscale image at position Pixel value at; : Perform a two-dimensional summation operation on elements within a 3×3 core range; S2.3. Combine the original intensity information of the image with gradient and texture features to construct a triple feature space of "intensity-gradient-texture", providing a comprehensive basis for subsequent similarity evaluation of matching blocks. S3 and PIMO projection iterative optimization search use four core operators (Residual Guided Projection RGP, Dual Random Projection DRP, Weighted Random Projection Update WRPU, and Lévy Flight Guided Projection LFGP) to find the optimal detail blocks, achieve accurate extraction of high-frequency details, and reconstruct the enhanced detail layer; Specifically, it includes: S3.1 Initialize the population and parameters: Randomly generate an initial population of size N, with the image space coordinates corresponding to the solution vectors, and the search dimension di; set the maximum number of iterations. Max iter Define the Jacobian matrix J For a 2×2 identity matrix, Lévy flight parameters β =1.5, β Initialize the elite solution for the shape parameters of the Lévy distribution. X best Used for iterative guidance to ensure the efficiency and stability of the optimization process; S3.2, Block Sampling and Matching Criterion Definition: The magnified image L1 is sampled in 5×5 blocks with a step size of 2 pixels. The intensity, gradient and texture triple features of each sampling block are extracted as reference blocks. The fitness function is defined as the sum of the absolute errors of the triple features of the reference block and the candidate block, which quantifies the matching similarity. ; in: Fitness score, used to evaluate the similarity of matched blocks; Perform a two-dimensional summation operation on the pixel features within a 5×5 block; : The reference block in the magnified image L1 Location intensity characteristics; Candidate blocks in smoothing layer L2 Location intensity characteristics; : The reference block in the magnified image L1 Gradient features of location; Candidate blocks in smoothing layer L2 Gradient features of location; : The reference block in the magnified image L1 Texture features of location; Candidate blocks in smoothing layer L2 Texture features of location; 0.001: Weighting coefficients for gradient and texture features; S3.3, Operator iteration execution, including: ① Residual-guided projection: Reference individuals are adaptively selected based on residual magnitude, and position updates are achieved through stochastic gradient calculation and Jacobian projection. The dynamic decay step size δ employs a sinusoidal decay strategy to balance exploration and development. ; in: Dynamic decay step size, used to control the update magnitude; : Current iteration number; Maximum number of iterations; : Normalized iteration progress, ranging from 0 to 2; : The 5th power of the iteration progress, used to adjust the decay rate.

[0033] Choose a strategy: ; in: : The solution vectors of the two selected reference individuals; : Probability of choice; : The fitness value of the i-th individual; The sum of the squares of the fitness values ​​of all individuals in the population; : A random number in the interval [0,1].

[0034] Gradient calculation: ; in: The calculated gradient vector; : Random numbers in the interval [0,1]; Reference Individual The difference vector between the solution and the elite solution; Reference Individual The difference vector between the solution and the elite solution; : A random number in the interval [0,1].

[0035] Location update: ; in: : The solution vector for the (t+1)th iteration after the projection operation; : The solution vector of the i-th individual in the current iteration; : The product of step size and gradient; : The product of the Jacobian matrix, step size, and gradient transpose.

[0036] ②Double random projection: Randomly select two individuals in the population as reference points, and generate update directions through two gradient combination methods to enhance global exploration capabilities; Randomly select two unique individual indices from the population. , Calculate two gradient vectors: ; ; in: : Gradient vectors of the two combinations; : Index is , The individual solution vector; The meanings of other symbols are the same as those for gradient calculation in the RGP operator.

[0037] Location update: ; in: : The solution vector after the (t+1)th iteration of the DRP operation, where "p2" is the operator index.

[0038] ③ Weighted random projection update: Two update paths are designed, and the random weights and iteration progress are adaptively selected to balance the breadth of exploration and the accuracy of development; Path 1 (Random Weighted Projection): ; in: : The solution vector obtained from the (t+1)th iteration using path 1; Random numbers (weighting coefficients) in the interval [0,1]. The solution vector for the current iteration; : The difference vector between the current solution and the elite solution.

[0039] Path 2: ; in: Adaptive coefficient, calculated using the following formula: ; : The solution vector obtained by the RGP operator.

[0040] The path selection rule is as follows: if random number 1 ÷ j > random number 2, then choose path 1; otherwise, choose path 2. "j" is an adjustment parameter, and its value is the population size N.

[0041] ④ Levy Flight Guidance Projection: Dynamically Adjusting Trigger Probability Based on Hyperbolic Tangent Function O By using Levy's flight to generate long-distance jump steps, the algorithm can be helped to find local optima. ; in: : The probability of triggering the LFGP operator, ranging from 0 to 1; Levi step size generation: ; in: Levi's stride length during flight; : Follows a pattern with a mean of 0 and a variance of Normally distributed random numbers; : Random numbers that follow a normal distribution with a mean of 0 and a variance of 1; : 1 / β power of the absolute value of v, where β is the shape parameter of the Lévy distribution.

[0042] The formula for calculating (the variance of u) is: ; in: Gamma function; Sine function operations.

[0043] Location update: ; in: The LFGP operator yields the first... The solution vector for the next iteration; Random weights in the interval [0,1]; The elite solution in the current iteration; Step size attenuation coefficient, calculated using the following formula: ; : Linear scaling factor for iteration progress.

[0044] S3.4 Boundary Processing and Elite Preservation: A reflection boundary strategy is used to limit the solution space range, ensuring that the updated position is within the effective image area. After each iteration, the individual with the best fitness is selected as the elite solution. X best This information is then passed on to the next iteration to ensure the correctness of the optimization direction. The reflection boundary strategy is as follows: If the updated position is lower than the lower bound, the reflection position is calculated using "2 × lower bound - updated position"; If the updated position is higher than the upper bound, the reflection position is calculated using "2 × upper bound - updated position"; If it is within the valid range, then maintain the original position; Wherein, the lower bound and the upper bound are the minimum and maximum valid values ​​of the image coordinates, respectively; S3.5, Detail layer reconstruction: The detail information corresponding to the best matching block obtained by all iterative search is accumulated and averaged, and the region with a count of zero is assigned a value of 1 to obtain the enhanced detail layer Residual.

[0045] S4. Detail fusion and enhancement overlay: The reconstructed detail layer is adapted and overlaid with the original image. The detail intensity is adjusted by the enhancement factor, and the pixel values ​​are cropped to ensure the image is reasonable, thus generating an enhanced image. Specifically, it includes: S4.1 Align the reconstructed detail layer Residual with bilinear interpolation according to the original image size to ensure spatial matching between detail information and the original image; S4.2 Perform enhanced overlay operations on each of the three RGB channels; the overlay formula is as follows: ; in: : Enhanced pixel value of channel c, O " represents output; "c" represents channel index, c=1,2,3 correspond to RGB channels; : The pixel value of the c-th channel of the original image; Enhancement factor; : The c-th channel pixel value of the aligned detail layer; S4.3. The superimposed pixel values ​​are truncated and limited to the effective pixel range of 0-255, and finally converted into an 8-bit unsigned integer format to obtain the enhanced image.

[0046] S5. Quality assessment and result output: Quantitatively evaluate the enhancement effect and export the processing results. Convert the enhanced image to uint8 type, calculate PSNR and SSIM indicators to evaluate the enhancement effect, and finally output the enhanced image.

[0047] Specifically, it includes: S5.1 Quantitative evaluation indicators are calculated using the Y channel of the YCbCr color space; the formulas for calculating peak signal-to-noise ratio and structural similarity are as follows: ; in: Peak signal-to-noise ratio; : Maximum pixel value, for 8-bit images ; Mean squared error, calculated using the following formula: .

[0048] ;

[0049] in: Structural similarity; : The average brightness value of the Y channel of the original image; Enhance the average brightness of the Y channel in the image; : Variance of the Y channel of the original image; Enhance the variance of the Y channel of the image; : Covariance of the Y channel between the original image and the enhanced image; : Adjustment constant, take , .

[0050] S5.2 supports manual saving of single images and automatic export of batch datasets. During batch processing, an "Enhanced Images" subfolder is generated to store all enhanced image files. At the same time, index statistics files in text and data formats are generated to record the PSNR, SSIM values ​​and overall average index of each image.

[0051] The algorithm framework of this invention is as follows: Figure 2 As shown, the algorithm consists of three modules: an image preprocessing module, a PIMO optimization detail extraction module, and an image enhancement and fusion module.

[0052] 1) The image preprocessing module determines the dimension of the input image. If it is a 3-channel color image, it is converted into a single-channel grayscale image using the rgb2gray function. If it is already a grayscale image, it is directly retained and converted into a double type to avoid numerical overflow or negative value truncation problems in subsequent convolution, gradient calculation and other operations.

[0053] 2) The PIMO optimization module obtains an intermediate image L1 by scaling the original image to 1.25 times, and then scales L1 back to its original size to obtain the base layer L2. The initial high-frequency detail layer H0 is obtained by calculating the difference between the original image and L2. The replicate mode is used to fill the boundaries of L2, L1, and H0, with the fill size being the search window size to avoid distortion in detail extraction at boundary regions. Gradient features are calculated using the grad function, and texture features are calculated using the texture function, resulting in gradient maps and texture maps for the base layer and intermediate layer, respectively, providing feature basis for matching search.

[0054] 3) The image enhancement and fusion module scales the detail image extracted by PIMO to the original image size using bilinear interpolation to ensure that the details are consistent with the dimensions of the original image. It sets an enhancement factor (default 2) and weights and superimposes the detail image and each channel of the original image to enhance the details such as image edges and textures. The pixel values ​​of the superimposed image are limited to the range of 0-255 to avoid image distortion caused by pixel overflow. The enhanced image is converted to uint8 type to meet the general requirements of image storage and display.

[0055] To intuitively verify the performance of the proposed PIMO algorithm under different enhancement factors, this invention provides objective numerical evaluation metrics on three classic datasets: RealSRet, BSD200, and T91. To objectively assess image enhancement quality, we selected the internationally accepted Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM) as evaluation standards. PSNR measures the noise and distortion of an image; a higher value indicates less distortion. SSIM assesses the structural fidelity of the image; a value closer to 1 indicates better structure preservation. Combining both metrics provides a more comprehensive reflection of image quality.

[0056] The PSNR is calculated as follows: ; Where MAX is the maximum value of the image pixels (usually 255), and MSE is the mean square error.

[0057] SSIM is calculated as follows: ; in, and These are the average brightness values ​​of the original image and the enhanced image, respectively. and The variances of the original image and the enhanced image are respectively. It is the covariance between the original image and the enhanced image. and It is a constant term, usually taken as =6.5025, =58.5225.

[0058] To intuitively verify the subjective enhancement effect of the algorithm of this invention, such as Figure 3 As shown, actual monitoring images of the coal drop point at the head of a mine conveyor belt were selected as test samples. This scenario has typical characteristics of low light and dust interference in underground mining, with equipment structural details and dust noise highly intertwined, placing stringent requirements on the algorithm's detail-noise balance capabilities.

[0059] As can be seen from the enhancement effect, traditional algorithms may suffer from problems such as blurred details, excessive noise amplification, and even artifacts after processing. However, the algorithm of this invention can accurately enhance details such as the outline of the tape support and the mesh of the protective net, while suppressing dust noise, making the equipment structure clear, the subjective visual effect natural, and the detail fidelity high, which is suitable for the image monitoring needs in complex mining scenarios.

[0060] To further illustrate the effectiveness of the algorithm, this invention also provides quantitative comparison results of PSNR and SSIM metrics with other algorithms on three classic datasets: BSD200, RealSRSet, and T91. The algorithms compared include GIF, GGIF, WGIF, BFLS, ILS, TH, QWLS, PMN, and NCC-PLM. Figure 4 and Figure 5 As shown. Figure 4 and Figure 5 The values ​​in bold are the maximum values, and the values ​​in underlined text are the suboptimal values.

[0061] This invention proposes an image detail enhancement algorithm based on PIMO projection iterative optimization, integrating multi-scale feature extraction and metaheuristic optimization strategies. Through the coordinated iteration of four projection operators, it achieves detail enhancement and noise suppression, achieving superior enhancement results compared to traditional algorithms without requiring large-scale labeled data and consuming low computational resources. It also possesses multi-scenario adaptability and real-time processing capabilities. This algorithm provides efficient and reliable image enhancement technology support for core scenarios such as underground monitoring, equipment diagnosis, and safety management in smart mines. It also offers a new technical path for image processing needs in fields such as industrial precision inspection and environmental monitoring, demonstrating significant technological innovation and promising industrial application prospects.

Claims

1. An image detail enhancement algorithm based on PIMO projection iterative optimization, characterized in that, Includes the following steps: S1. Image preprocessing and multi-scale decomposition; S2. Multi-dimensional feature calculation constructs a triple feature space of "intensity-gradient-texture" to provide a basis for matching evaluation; S3 and PIMO projection iterative optimization search find the optimal detail block through operators and reconstruct the enhanced detail layer; S4. Detail Blending and Enhancement Overlay: The reconstructed detail layer is adapted and overlaid with the original image to generate an enhanced image; S5. Quality assessment and result output: Quantitative evaluation enhances the effect and exports the processing results.

2. An image detail enhancement algorithm based on PIMO projection iterative optimization according to claim 1, characterized in that, S1 image preprocessing and multi-scale decomposition specifically include: S1.1 Perform format standardization processing on the input image; S1.

2. Multi-scale scaling is performed through bilinear interpolation to generate a magnified image L1 and a smoothing layer L2 restored to the original size. The difference between the original image and the smoothing layer L2 is calculated to obtain the initial detail layer H0, thus achieving smoothing-detail separation of the image. S1.

3. Use the copy-fill method to expand the boundaries of images at each scale to avoid artifacts in the edge region during feature calculation and block matching.

3. An image detail enhancement algorithm based on PIMO projection iterative optimization according to claim 2, characterized in that, S2 multi-dimensional feature calculation specifically includes: S2.1 Calculate the horizontal and vertical gradients of smoothing layers at each scale based on the Sobel operator, and integrate the bidirectional gradient information through the gradient magnitude formula to construct gradient features; S2.

2. Perform convolution operation on the grayscale image using a Laplacian convolution kernel to generate a texture response map and extract texture features; S2.3 Combine the original intensity information of the image with gradient and texture features to construct a triple feature space of "intensity-gradient-texture", which provides a comprehensive basis for the similarity evaluation of subsequent matching blocks.

4. An image detail enhancement algorithm based on PIMO projection iterative optimization according to claim 3, characterized in that, S3PIMO projection iterative optimization search specifically includes: S3.1 Initialize the population and parameters: Randomly generate an initial population of size N, with the image space coordinates corresponding to the solution vectors, and the search dimension di; set the maximum number of iterations. Maxiter Define the Jacobian matrix J Levi's flight parameters β Initialize elite solution X best Used for iterative guidance; S3.2, Block Sampling and Matching Criterion Definition: The magnified image L1 is sampled in 5×5 blocks with a step size of 2 pixels. The intensity, gradient and texture triple features of each sampling block are extracted as reference blocks. The fitness function is defined as the sum of the absolute errors of the triple features of the reference block and the candidate block, which quantifies the matching similarity. S3.3, Operator iteration execution, including: ① Residual-guided projection: Reference individuals are adaptively selected based on the residual size, and position updates are achieved through stochastic gradient calculation and Jacobian projection; ②Double random projection: Randomly select two individuals in the population as reference points, and generate update directions through two gradient combination methods to enhance global exploration capabilities; ③ Weighted random projection update: Two update paths are designed, and the random weights and iteration progress are adaptively selected to balance the breadth of exploration and the accuracy of development; ④ Levy Flight Guidance Projection: Dynamically Adjusting Trigger Probability Based on Hyperbolic Tangent Function O By using Levy's flight to generate long-distance jump steps, the algorithm can be helped to find local optima. S3.4 Boundary Processing and Elite Preservation: A reflection boundary strategy is used to limit the solution space range, ensuring that the updated position is within the effective image area. After each iteration, the individual with the best fitness is selected as the elite solution. X best This information is then passed on to the next iteration to ensure the correctness of the optimization direction. S3.5, Detail layer reconstruction: The detail information corresponding to the best matching block obtained by all iterative search is accumulated and averaged, and the region with a count of zero is assigned a value of 1 to obtain the enhanced detail layer Residual.

5. An image detail enhancement algorithm based on PIMO projection iterative optimization according to claim 4, characterized in that, The reflection boundary strategy is as follows: If the updated position is lower than the lower bound, the reflection position is calculated using "2 × lower bound - updated position"; If the updated position is higher than the upper bound, the reflection position is calculated using "2 × upper bound - updated position"; If it is within the valid range, then maintain the original position; The lower and upper bounds represent the minimum and maximum valid values ​​of the image coordinates, respectively.

6. An image detail enhancement algorithm based on PIMO projection iterative optimization according to claim 4 or 5, characterized in that, S4 Detail Blending and Enhanced Overlay specifically includes: S4.1 Align the reconstructed detail layer Residual with bilinear interpolation according to the original image size to ensure spatial matching between detail information and the original image; S4.2 Perform enhancement overlay operations on the three RGB channels separately; S4.

3. The superimposed pixel values ​​are truncated and limited to the effective pixel range of 0-255, and finally converted into an 8-bit unsigned integer format to obtain the enhanced image.

7. An image detail enhancement algorithm based on PIMO projection iterative optimization according to claim 6, characterized in that, S5 quality assessment and results output specifically include: S5.1 Calculate the quantitative evaluation index using the Y channel of the YCbCr color space; S5.2 supports manual saving of single images and automatic export of batch datasets. During batch processing, an "Enhanced Images" subfolder is generated to store all enhanced image files. At the same time, index statistics files in text and data formats are generated to record the PSNR, SSIM values ​​and overall average index of each image.

Images