A perception-guided UAV image dehazing algorithm based on superpixel scene prior

Abstract

The present invention provides a defogging algorithm, which introduces superpixel scene prior to reduce the calculation of the UAV defogging process. The image defogging process can be converted from RGB color space to Lab color space to avoid confusion and redundant information between different chromaticities, so as to select reliable areas for efficient defogging through the L channel. Considering the influence of light inhomogeneity in the defogging process, the present invention designs a guided filtering algorithm based on simple linear iterative clustering, which replaces a large number of guided windows with superpixel clustering windows with similar color blocks while retaining complementary information. In order to improve the perception ability after defogging, the superpixel segmentation and target detection results are quantitatively analyzed, and the alternating direction multiplier method is used to design a collaborative optimization feedback iteration mechanism of perception and defogging to enhance the effectiveness of UAV visual tasks in foggy environments. The framework proposed by the present invention can effectively defog UAV images and significantly improve their perception processing capabilities.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a perception-guided UAV image defogging algorithm based on super-pixel scene prior. Background Art

[0002] Aerial imaging systems carried by drones have been applied in many fields, such as urban monitoring, power and oil pipeline inspection, agricultural and forestry operations, etc., and are of great significance for performing perception tasks in these applications. However, due to climate change, air pollution and other reasons, it is inevitable that drones will shoot foggy images. In foggy scenes, suspended particulate matter, such as water vapor and dust, will cause the image quality to be significantly reduced. During the propagation process, air particles will scatter and absorb light, resulting in a decrease in the intensity of the reflected light of the target object. At the same time, the surrounding atmospheric light will also be scattered by air particles and become part of the imaging. Therefore, the final image will inevitably show the phenomenon of color attenuation and reduced contrast. In the images obtained by drones, detailed information is seriously lost and the blurriness increases, which will reduce the visual effect to varying degrees and affect the perception function of the aerial monitoring system. Therefore, a defogging method is needed to preprocess such low-quality images to ensure the normal operation of the perception system.

[0003] Earlier research works were based on basic image processing, attempting to improve the clarity of images by adjusting contrast, brightness, and saturation locally or globally on the image. However, these methods cannot essentially solve the problem of image haze, and these simple operations may only play a limited role in the image dehazing process. In essence, many researchers have exploited prior information from physical models for dehazing. Various analytical and statistical methods have been used in different fields to achieve image dehazing results, but the performance still needs to be improved. In recent years, deep learning-based image dehazing methods have been proposed, which can be broadly classified into parameter generation methods and direct synthesis methods. Although previous studies have achieved promising results by highlighting areas that are important for image dehazing, they have focused too much on image dehazing itself, thus ignoring the extent to which image dehazing affects downstream tasks. Summary of the invention

[0004] The purpose of the present invention is to provide a UAV image defogging algorithm that is oriented to perception tasks, starts from the basic texture features and perception features in the defogging process, and uses the idea of ​​collaborative optimization to improve the defogging and perception effects in the iterative feedback process, so as to solve at least one technical problem existing in the above-mentioned background technology. In the first aspect, a perception-oriented UAV image defogging algorithm based on superpixel scene prior is provided, comprising the following steps:

[0005] From multiple drone images in a sequence, one image is randomly selected as the processing image;

[0006] Process the image according to the super-pixel scene prior method, use the filtering algorithm to perform preliminary dehazing on the image and improve the possible white area impact;

[0007] Use the SLIC algorithm for superpixel segmentation, and use the superpixel blocks and superpixel images generated during the processing as guidance windows and guidance maps for filtering;

[0008] Calculate the atmospheric illumination intensity and perform dehazing on a single image;

[0009] The superpixel and target recognition feature parameters before and after defogging are extracted and quantified to determine whether the optimization conditions of the alternating direction multiplier method are met: if so, the atmospheric illumination intensity is recalculated and the defogging process of the single image is iteratively optimized; if not, the optimal defogging parameters are obtained.

[0010] Based on the above solution, the superpixel scene prior method specifically includes:

[0011] After the processed image is segmented by superpixels, the chrominance components a and b are used to remove fog from the image;

[0012] Define the mathematical model of the gray component G of the fog image and obtain the corresponding color component;

[0013] According to the linear relationship between RGB and Lab color spaces, the gray component G is projected into the L channel of the Lab color space to obtain the precise turbidity feature area.

[0014] Based on the above solution, the filtering algorithm is specifically an improved uniform filtering algorithm, which specifically includes:

[0015] The image function g(x,y) is expressed as the product of the illumination function i(x,y) and the reflection function r(x,y). The image model and its value range are as follows:

[0016] ;

[0017] ;

[0018] Take the logarithm of the image function g(x, y), perform Fourier transform to the frequency domain, select the filter function to control the components on the right, and obtain the filtering result:

[0019] ;

[0020] Among them, G(u,v), I(u,v), R(u,v) are the frequency domain functions of g(x,y), i(x,y), r(x,y) respectively; S(u,v) is the filtering function:

[0021] ;

[0022] Where C(u,v) is the distance from the frequency point (u,v) to the center point of the filter; 0.7≤α≤1.4, 0.01≤β≤0.5.

[0023] Based on the above scheme, the guided filtering expression for filtering the guided window and the guided graph is:

[0024] ;

[0025] ;

[0026] Where q represents the output image; Indicates the guiding image; represents the input image; r is the window radius; s is the window pixel; is the regularization coefficient; and is the window constant coefficient; represents the filter window, The constraint equation is:

[0027] ;

[0028] Take the minimum value of the constraint equation to get the window constant coefficient and It is expressed as follows:

[0029] ;

[0030] ;

[0031] in, and The windows are Middle guide map The mean and variance of and The windows are Input image The mean and total number of elements of ;

[0032] The filter window will be guided Replace superpixel blocks near the cluster center in SLIC in a small local area;

[0033] Use superpixel blocks in the SLIC algorithm to calculate traditional coefficients and To reduce the feature dimension of SLIC:

[0034] ;

[0035] ;

[0036] ;

[0037] in: represents the superpixel block of SLIC; D is the distance between each pixel and the cluster center; is the grayscale feature distance; is the spatial characteristic distance; m is the spatial distance influence factor, is the distance between seed points, and are the grayscale color space coordinates of points j and h respectively.

[0038] Based on the above scheme, the atmospheric light intensity algorithm is specifically as follows:

[0039] ;

[0040] ;

[0041] ;

[0042] ；

[0043] in, and They are the processing results of the grayscale morphological operation results F(x), Q(x) and T(x); Indicates that the atmospheric light intensity parameter A is within the maximum value of the original image pixel; λ is the adaptive threshold, with a value range of [0,1], indicating that The average value of all elements of .

[0044] Based on the above scheme, for each pixel in the superpixel, the grayscale value difference is compared with that of the pixels in eight directions. If the grayscale value of the pixel in other directions is greater than the grayscale value of the central pixel, it is set to 1, otherwise it is set to 0.

[0045] Finally, each pixel has a corresponding eight-bit binary number, which will generate 256 eigenvalues, which are converted into decimal values ​​as the LBP eigenvalues ​​of the pixel. The specific steps of extracting and quantifying the superpixel and target recognition feature parameters before and after defogging are as follows:

[0046] Compute texture features of superpixels:

[0047] ;

[0048] ;

[0049] Where N is the superpixel block The number of pixels in is the LBP feature value of each pixel, and w is the superpixel block in the entire image The number of It is the superpixel texture feature of the image;

[0050] Use yolov7 algorithm and train based on visdrone dataset;

[0051] Use the trained model to perform target detection on the images before and after dehazing, and record the precision P, recall rate R and confidence R of the model's recognition results for the detection set, the recall rate R and the weighted sum average F, and the detection time T;

[0052] By and Combined in negative feedback form, the following equation is constructed:

[0053] ;

[0054] ;

[0055] in, is the introduced free variation vector.

[0056] On the basis of the above scheme, When the F2 score is used as the final accuracy evaluation of the model effect, the formula for calculating F is:

[0057] .

[0058] On the basis of the above scheme, the specific algorithm for judging whether the optimization conditions of the alternating direction multiplier method are met is:

[0059] The augmented Lagrange function of the objective function implemented by the ADMM algorithm is:

[0060] ;

[0061] Among them, z is the Lagrange multiplier, λ>0 is the penalty parameter;

[0062] For the above objective function, the ADMM algorithm is as follows:

[0063] ;

[0064] ;

[0065] ;

[0066] Use in multiplication update As the step size, where:

[0067] ;

[0068] ;

[0069] ;

[0070] The following inequality is satisfied:

[0071] ;

[0072] in, is a set of accuracy thresholds, It is an adjustable parameter;

[0073] ;

[0074] ;

[0075] in, s is the dual residual, r is the primary residual, and k is the number of iterations.

[0076] In a second aspect, a computer device is provided, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any of the above methods when executing the computer program.

[0077] According to a third aspect, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the steps of any of the above methods are implemented.

[0078] Beneficial effects of the present invention:

[0079] The present invention proposes a UAV image defogging framework for solving perception tasks under foggy conditions. Different from pixel-level defogging, a superpixel scene prior is introduced to reduce the calculation of the UAV defogging process. The image defogging process can be converted from the RGB color space to the Lab color space, which can avoid confusion and redundant information between different chromaticities, so that reliable areas for efficient defogging can be selected through the L channel. Considering the influence of light inhomogeneity in the defogging process, the present invention uses a guided filtering algorithm based on SLIC, which replaces a large number of guiding windows with superpixel clustering windows with similar color blocks while retaining complementary information. At the same time, under the guidance of SPSP, SLIC is also processed in the Lab color space, so that the defogging speed can be improved. In order to improve the perception ability after defogging, the superpixel segmentation and target detection results are quantitatively analyzed, and the alternating direction multiplier method (ADMM) is used to design a collaborative optimization feedback iteration mechanism of perception and defogging to enhance the effectiveness of UAV visual tasks in foggy environments. The present invention has a better effect than previous methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] The present invention has the following accompanying drawings:

[0081] Figure 1 The present invention is a flowchart of a UAV image dehazing algorithm based on superpixel scene prior perception guidance according to a specific embodiment of the present invention. DETAILED DESCRIPTION

[0082] To make the objects, advantages and features of the present invention more apparent, the following Figure 1 The present invention is further described in detail with reference to the following specific embodiments.

[0083] Embodiment 1 of the present invention provides a superpixel scene prior-based perception-guided UAV image defogging algorithm, which includes the following process steps:

[0084] Step S110: Randomly select an image from multiple drone images in a sequence as a processing image.

[0085] Step S120: According to the superpixel scene prior, the obtained image is transferred from the RGB color space to the Lab color space and calculated. The fog is concentrated on the brightness component L of the Lab color space and processed. The image is preliminarily defogged using an improved homogeneous filtering algorithm and the possible influence of white areas is improved.

[0086] Step S130: Use the SLIC-based guided filtering algorithm. First, use the SLIC algorithm to perform superpixel segmentation, and use the superpixel blocks and superpixel images generated during the processing as the guiding window and guiding map for filtering.

[0087] Step S140: Calculate the atmospheric illumination intensity A and perform defogging on the single image.

[0088] Step S150: extract and quantify the superpixels and target recognition feature parameters before and after defogging, and determine whether the optimization conditions of the alternating direction multiplier method are met. If so, re-execute step S150 for iterative optimization; if not, obtain the optimal defogging parameters.

[0089] Step S160: Apply the optimal defogging parameters guided by the perception of a single image to the sequence of images to complete the defogging work of the UAV in the mission area.

[0090] In step S120, the image in the RGB space must first be converted into the image in the Lab space. First, the channel must be converted into the XYZ space. The specific conversion relationship equation is as follows:

[0091] in is a 3×3 matrix: After that, the image is converted from XYZ space to Lab space. The specific conversion relationship is as follows:

[0092] ;

[0093] ;

[0094] ;

[0095] .

[0096] After conversion to Lab color space, the improved homomorphic filtering algorithm is used for filtering:

[0097] ;

[0098] Where: C(u,v) is the distance from the frequency point (u,v) to the center point of the filter; α and β range from 0.7≤α≤1.4 and 0.01≤β≤0.5. In a preferred embodiment, the improved uniform filtering algorithm takes α=0.97 and β=0.28.

[0099] The image is then sliced ​​into superpixel blocks: 1) Divide the grid according to the image size and initialize the cluster center; 2) Move the cluster center to the minimum gradient position in the adjacent interval; 3) Set the best matching pixel in the spatial adjacent area near each cluster center through the distance measurement formula; 4) Calculate the error of the new cluster center and force it to connect after meeting the threshold. The distance measurement involved is as follows:

[0100] ;

[0101] ;

[0102] ;

[0103] Among them, J is the distance between each pixel and the cluster center; is the color feature distance; and are the Lab color space coordinates of grid pixels j and h respectively; is the spatial feature distance; and are the two-dimensional spatial coordinates of points j and k respectively; m is the spatial distance influence factor; τ is the distance between seed points.

[0104] The guided filtering expression is as follows:

[0105] ;

[0106] ;

[0107] Where q represents the output image; Indicates the guiding image; represents the input image; r is the window radius; s is the window pixel; is the regularization coefficient; and is the window constant coefficient; represents the filter window, The constraint equation is:

[0108] ;

[0109] The window constant coefficient is obtained by minimizing the deviation between the input and output images by taking the minimum value of the constraint equation and It is expressed as follows:

[0110] ;

[0111] ;

[0112] in, and The windows are Middle guide map The mean and variance of and The windows are Input image The mean and total number of elements of ;

[0113] Guided filter window In a small local area, the superpixel blocks near the cluster center in SLIC can be substituted with each other. In the present invention, the superpixel blocks in the SLIC algorithm are used to calculate (43) and (44) to reduce the feature dimension of SLIC, as shown in (45)-(47).

[0114] ;

[0115] ;

[0116] ;

[0117] in: represents the superpixel block of SLIC; D is the distance between each pixel and the cluster center; is the grayscale feature distance; is the spatial characteristic distance; m is the spatial distance influence factor, is the distance between seed points, and are the grayscale color space coordinates of points j and h respectively.

[0118] Then, the superpixel blocks and superpixel images generated in the processing are used as guide windows and guide maps for filtering. The adaptive adjustment calculation of the illumination intensity A is realized using formulas (48)-(50); finally, the obtained parameters are substituted into formula (51) to complete the restoration of the fog image.

[0119] ;

[0120] ;

[0121] ;

[0122] ;

[0123] Using SLIC superpixel segmentation and traditional target detection algorithms, comprehensive characterization extraction and constraint setting before and after defogging of drone images are implemented, and appropriate defogging parameters are selected through iterative optimization of a single image. The texture is represented using the local binary pattern (LBP) algorithm, where the LBP algorithm needs to be extracted within a rectangular or circular area. Therefore, in this invention, we use the outer rectangle of each superpixel and calculate the outer part based on the grayscale value of the center point of the superpixel. For each pixel in the superpixel, it is compared with the grayscale value difference of the pixels in eight directions. If the grayscale value of the pixel in other directions is greater than the grayscale value of the central pixel, it is set to 1, otherwise it is set to 0.

[0124] Each pixel has a corresponding eight-bit binary number, which will generate 256 eigenvalues, which are converted into decimal values ​​as the LBP eigenvalues ​​of the pixel. The formula for calculating the texture features of superpixels is as follows:

[0125] ;

[0126] ;

[0127] Where N is the superpixel block The number of pixels in is the LBP feature value of each pixel, and w is the superpixel block in the entire image The number of It is the super-pixel texture feature of the image.

[0128] The original yolov7 algorithm was used, and 1,000 foggy images and 1,000 defogging images were added for training based on the visdrone dataset. The trained model was then used to perform target detection on drone images before and after defogging. The precision P (Precision), recall R (Recall) and confidence R (Confidence) of the model's recognition results for the detection set were recorded, as well as the recall R (Recall) and weighted sum average F, and the detection time T. The F2 score was used as the final accuracy evaluation of the model's effect.

[0129] After obtaining the super-pixel texture features After the perception-oriented feature representation before and after dehazing and the composite feature F2 of target detection, the following equation is constructed:

[0130] ;

[0131] ;

[0132] The superpixel texture features and comprehensive features of target detection will be improved. The greater the improvement, the better the dehazing effect. Therefore, the following objective function can be constructed:

[0133] ;

[0134] Using ADMM to solve the above equation, the augmented Lagrange function of the objective function is:

[0135] ;

[0136] For the above objective function, the ADMM algorithm is as follows:

[0137] ;

[0138] ;

[0139] ;

[0140] We use τ as the step size in the multiplicative update, where:

[0141] ;

[0142] ;

[0143] ;

[0144] Equation (60) is the pairwise residual, (61) is the principal residual, and the following inequalities are satisfied:

[0145] ;

[0146] in is a set of accuracy thresholds, and It is an adjustable parameter.

[0147] ;

[0148] ;

[0149] When the iterative convergence conditions are met, the optimal dehazing parameters for a single image can be obtained, thus completing the perception-oriented dehazing optimization algorithm.

[0150] The optimal defogging parameters guided by single image perception are applied to the sequence of images to complete the defogging work of the UAV in the mission area.

[0151] Comparison of experimental results:

[0152] In this Example 1, a perception-guided drone image defogging algorithm based on superpixel scene prior is compared with six existing image defogging algorithms, and the experimental data used is collected by itself. In March 2022, December 2022, and March 2023, multi-height and multi-angle collections were carried out using a DJI Mavic 2 Enterprise Advanced drone in urban and suburban locations in Shijiazhuang, Hebei Province, China, when sandstorms occurred. The collection time includes early morning, afternoon, and evening, and the targets in the images include power line towers, cars, vans, trucks, excavators, etc. The image resolution is 1920×1080, and there are 3000 images in total.

[0153] Table 1 shows the comparison between the present invention and six existing image defogging algorithms in terms of contrast Γ value, average gradient G value, color deviation K value and information entropy Λ value. The relevant calculation formulas are as follows:

[0154] ;

[0155] ;

[0156] ;

[0157] ;

[0158] stands for contrast, which characterizes the ability to express image details; L(x) represents the brightness of pixel x; N is the number of pixels; G represents the average gradient, where h, w, They represent the height, width, horizontal gradient, and vertical gradient of the image respectively; K represents the color deviation, which characterizes the degree of color deviation; , Respectively represent the means of a and b in Lab space; is the mean square error of L in Lab space; represents the deviation threshold; stands for information entropy, which characterizes the richness of image information; H is the number of gray levels; p(e) is the probability of gray level distribution.

[0159] Table 1

[0160]

[0161] Among them, C, D, E, F, G, and H represent HEkaiming algorithm, D.Wierzbicki algorithm, H.Yuqing algorithm, S.Wang, Y.Li algorithm, and R.Zheng algorithm respectively.

[0162] It can be seen from Table 1 that, compared with other algorithms, the algorithm of the present invention significantly improves the contrast of the original image and effectively reduces the degree of color deviation; at the same time, the algorithm of the present invention has the largest information entropy value, which indicates that it can significantly remove the fog of the power inspection drone image and restore a large amount of detailed information in the image.

[0163] Embodiment 2 of the present invention provides a computer device, including a memory and a processor, the processor and the memory communicate with each other, the memory stores program instructions executable by the processor, and the processor calls the program instructions to execute a super-pixel scene prior-based perception-guided drone image defogging algorithm, including the following process steps:

[0164] Step S110: Randomly select an image from multiple drone images in a sequence as a processing image.

[0165] Step S120: According to the superpixel scene prior, the obtained image is transferred from the RGB color space to the Lab color space and calculated. The fog is concentrated on the brightness component L of the Lab color space and processed. The image is preliminarily defogged using an improved homogeneous filtering algorithm and the possible influence of white areas is improved.

[0166] Step S130: Use the SLIC-based guided filtering algorithm. First, use the SLIC algorithm to perform superpixel segmentation, and use the superpixel blocks and superpixel images generated during the processing as the guiding window and guiding map for filtering.

[0167] Step S140: Calculate the atmospheric illumination intensity A and perform defogging on the single image.

[0168] Step S150: extract and quantify the superpixels and target recognition feature parameters before and after defogging, and determine whether the optimization conditions of the alternating direction multiplier method are met. If so, re-execute step S150 for iterative optimization; if not, obtain the optimal defogging parameters.

[0169] Step S160: Apply the optimal defogging parameters guided by the perception of a single image to the sequence of images to complete the defogging work of the UAV in the mission area.

[0170] Embodiment 3 of the present invention provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, a perception-guided UAV image defogging algorithm based on superpixel scene prior is implemented, including the following process steps:

[0171] Step S110: Randomly select an image from multiple drone images in a sequence as a processing image.

[0172] Step S120: According to the superpixel scene prior, the obtained image is transferred from the RGB color space to the Lab color space and calculated. The fog is concentrated on the brightness component L of the Lab color space and processed. The image is preliminarily defogged using an improved homogeneous filtering algorithm and the possible influence of white areas is improved.

[0173] Step S130: Use the SLIC-based guided filtering algorithm. First, use the SLIC algorithm to perform superpixel segmentation, and use the superpixel blocks and superpixel images generated during the processing as the guiding window and guiding map for filtering.

[0174] Step S140: Calculate the atmospheric illumination intensity A and perform defogging on the single image.

[0175] Step S150: extract and quantify the superpixels and target recognition feature parameters before and after defogging, and determine whether the optimization conditions of the alternating direction multiplier method are met. If so, re-execute step S150 for iterative optimization; if not, obtain the optimal defogging parameters.

[0176] Step S160: Apply the optimal defogging parameters guided by the perception of a single image to the sequence of images to complete the defogging work of the UAV in the mission area.

[0177] The above implementation modes are only used to illustrate the present invention, but not to limit the present invention. Ordinary technicians in the relevant technical field can make various changes and modifications without departing from the essence and scope of the present invention. Therefore, all equivalent technical solutions also belong to the scope of the present invention, and the patent protection scope of the present invention should be defined by the claims. The contents not described in detail in this specification belong to the prior art known to professional and technical personnel in this field.

Claims

1. A perception-guided UAV image dehazing method based on superpixel scene prior, characterized in that: The steps include: From multiple images in a sequence, randomly select an image as a processing image; Process the image according to the super-pixel scene prior method, use the filtering algorithm to preliminarily remove the haze from the image and improve the possible white area impact; Use the SLIC algorithm for superpixel segmentation, and use the superpixel blocks and superpixel images generated during the processing as guidance windows and guidance maps for filtering; Calculate the atmospheric light intensity and perform dehazing on a single image; Extract and quantify the superpixel and target recognition feature parameters before and after defogging to determine whether the optimization conditions of the alternating direction multiplier method are met: if so, recalculate the atmospheric illumination intensity and perform iterative optimization of the defogging process on the single image; if not, obtain the optimal defogging parameters; After the processed image is segmented by superpixels, the chrominance components a and b are used to remove fog from the image; Define the mathematical model of the gray component G of the fog image and obtain the corresponding color component; According to the linear relationship between RGB and Lab color spaces, the gray component G is projected into the L channel of the Lab color space to obtain the precise turbidity feature area; The filtering algorithm is specifically an improved uniform filtering algorithm, which specifically includes: The image function g(x,y) is expressed as the product of the illumination function i(x,y) and the reflection function r(x,y). The image model and its value range are as follows: ； ； Take the logarithm of the image function g(x, y), perform Fourier transform to the frequency domain, select the filter function to control the components on the right, and obtain the filtering result: ； Among them, G(u,v), I(u,v), R(u,v) are the frequency domain functions of g(x,y), i(x,y), r(x,y) respectively; S(u,v) is the filtering function: ； Where C(u,v) is the distance from the frequency point (u,v) to the center point of the filter; 0.7≤α≤1.4, 0.01≤β≤0.5; The guided filter expression for filtering the guided window and the guided graph is: ； ； Where q represents the output image; I represents the guide image; φ represents the input image; r is the window radius; s is the window pixel; ε is the regularization coefficient; and v is the window constant coefficient; represents the filter window, The constraint equation is: ； Take the minimum value of the constraint equation to get the window constant coefficient and It is expressed as follows: ； ； in, and The windows are Middle guide map The mean and variance of and The windows are Input image The mean and total number of elements of ; The filter window will be guided Replace superpixel blocks near the cluster center in SLIC in a small local area; Use superpixel blocks in the SLIC algorithm to calculate traditional coefficients and To reduce the feature dimension of SLIC: ； ； ； in: represents the superpixel block of SLIC; D is the distance between each pixel and the cluster center; is the grayscale feature distance; is the spatial characteristic distance; m is the spatial distance influence factor, is the distance between seed points, and are the grayscale color space coordinates of points j and h respectively; The atmospheric light intensity algorithm is specifically: ； ； ； ； ； in, and They are the processing results of the grayscale morphological operation results F(x), Q(x) and T(x); Indicates that the atmospheric light intensity parameter A has a value range not greater than the maximum value of the original image pixel; λ is an adaptive threshold value, with a value range of [0,1], indicating that The average value of all elements of ; The specific steps of extracting and quantifying superpixels and target recognition feature parameters before and after defogging are as follows: Compute texture features of superpixels: ； ； Where N is the superpixel block The number of pixels in is the LBP feature value of each pixel, and w is the superpixel block in the entire image The number of It is the superpixel texture feature of the image; Use yolov7 algorithm and train based on visdrone dataset; Use the trained model to perform target detection on the images before and after dehazing, and record the precision P, recall rate R and confidence R of the model's recognition results for the detection set, the recall rate R and the weighted sum average F, and the detection time T; By and Combined in negative feedback form, the following equation is constructed: ； ； in, is the introduced free variation vector; Construct the objective function: 。 2. The defogging method according to claim 1, characterized in that: The formula for calculating F is: 。 3. The defogging method according to claim 2, characterized in that: The specific algorithm for determining whether the optimization conditions of the alternating direction multiplier method are met is: The augmented Lagrange function of the objective function implemented by the ADMM algorithm is: ； Among them, z is the Lagrange multiplier, λ>0 is the penalty parameter; For the above objective function, the ADMM algorithm is as follows: ； ； ； We use τ as the step size in the multiplicative update, where: ； ； ； The following inequality is satisfied: ； Among them, ε is a set of accuracy thresholds, and η is an adjustable parameter; ； ； in, s is the dual residual, r is the primary residual, and k is the number of iterations.

4. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 3 are implemented.

5. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 3 are implemented.

Images