Reservoir bank slope crack identification method based on unmanned aerial vehicle inspection and digital image

Abstract

The invention discloses a reservoir bank slope crack identification method based on unmanned aerial vehicle inspection and digital images, and relates to the technical field of reservoir bank slope inspection. 2, image preprocessing: preprocessing the data by adopting processing methods of graying, gray enhancement and bilateral filtering; 3, crack identification: identifying reservoir bank slope cracks in the image by adopting methods of threshold segmentation, edge detection and supervised classification; 4, crack pixel filtering: performing filtering processing on the extracted crack image by adopting morphological closed operation, length filtering and direction filtering; and 5, model fusion: fusing the threshold segmentation model, the edge detection model and the supervised classification model together by adopting a minimum-risk Bayesian probability method. The method shows remarkable applicability and effect in large-range and multi-form crack identification and investigation, and shows high robustness when processing crack images with different scales, shapes and backgrounds, and the accuracy of the method can stably reach more than 90%.

Description

Technical Field

[0001] The invention relates to the technical field of reservoir bank slope inspection, and in particular to a reservoir bank slope crack identification method based on unmanned aerial vehicle inspection and digital images. Background Art

[0002] As an indispensable and important part of my country's water conservancy construction, reservoirs have multiple functions, including flood control, power generation, irrigation, water supply, shipping, etc. However, reservoirs are often built in mountainous canyon areas. The geological structure of these areas is complex, the rock strata are varied, the rock properties are significantly different, and the mechanical properties and stability of the rocks are affected by many factors. When the reservoir is filled with water, the effect of water will affect the rocks and soil around the reservoir, causing the stability of the slope to be challenged and the formation of cracks to become a common phenomenon. However, most of the crack areas are located in the middle and upper parts of the mountains, with steep terrain. It is difficult to quickly and accurately detect and measure large-scale surface cracks by relying solely on manual on-site surveys.

[0003] In the existing technology, attention has begun to be paid to the use of drone photogrammetry technology to shoot high-resolution digital orthophotos, three-dimensional point clouds and digital surface models. These technologies can efficiently obtain surface information in a large area, making the identification of cracks faster and more accurate. Drones can conduct inspections on complex terrain, overcoming the limitations of manual inspections, especially when cracks are widely distributed, complex in shape and with a lot of background noise. However, at present, the application of drones still relies on the participation of professional technicians (including drone operators, professional geological engineers, etc.), and the entry threshold is high, making it difficult to become an effective part of the daily management of geological disasters in hydropower stations.

[0004] With the development of digital image processing technology, a large number of research results have been achieved in crack recognition in relatively simple backgrounds such as concrete pavements, tunnel linings, building walls, bridges, etc. However, in complex and diverse backgrounds, crack recognition using only a single algorithm model is particularly weak. Its main drawback is that in complex and diverse backgrounds, the shape, size, location and other characteristics of cracks vary greatly. A single algorithm model often cannot accurately capture these subtle differences and cannot handle crack recognition problems in various situations, resulting in challenges to the accuracy of crack recognition. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention proposes a method for more comprehensively capturing the characteristics of cracks in complex backgrounds by combining the complementary advantages of different algorithm models, thereby improving the accuracy and robustness of crack detection. At the same time, the method is combined with the drone inspection technology, and the drone inspection technology is used to automatically shoot and generate orthophoto images and automatically submit them to the program for crack identification, thereby truly realizing the full-process automated processing of the reservoir slope crack identification method based on drone inspection and digital images. The specific technical scheme is as follows: A reservoir slope crack identification method based on UAV inspection and digital images: The specific steps are: Step 1: Data acquisition, using drones to collect reservoir slope image data; Step 2: Image preprocessing, using grayscale, grayscale enhancement and bilateral filtering to process the image. Preprocess the data; Step 3: Crack identification: using threshold segmentation, edge detection and supervised classification methods to identify the cracks on the reservoir slope in the image; Step 4: Crack pixel filtering, using morphological closing operation, length filtering and direction filtering to filter the extracted crack image; Step 5: Model fusion, using the minimum risk Bayesian probability method to fuse the three models of threshold segmentation, edge detection and supervised classification.

[0006] As an optimization: the step 2 is specifically as follows; 2.1. According to the digital television signal standard, the original RGB image is averaged with a weight of red: green: blue = 0.298: 0.587: 0.114, and a single-channel grayscale image is generated from the three-channel RGB image to eliminate the additional complexity introduced by color information and highlight the texture, shape and structure information of the image; 2.2. Use the histogram equalization method to enhance the grayscale of the image. First, use the matrix to collect the entire image and identify the pixel distribution of each grayscale level. Finally, perform grayscale mapping based on the cumulative distribution function to assign each pixel value in the original image to a new grayscale level to achieve grayscale enhancement. 2.3. Perform bilateral filtering to denoise the image, including Gaussian noise and salt and pepper noise, so that the image will not be blurred or distorted while denoising. By selecting the size and standard deviation of the filter, the details and edge information of the image can be retained while removing noise.

[0007] As an optimization: the step three is specifically as follows; 3.1. Threshold segmentation is as follows: After the digital image collected by the drone is grayed, the grayscale threshold H is set, and the grayscale value less than H is set as the foreground, and the grayscale value greater than H is set as the background, so as to realize crack extraction; 3.2. Edge detection is a process of judging whether a pixel belongs to the edge of an object by the change in grayscale gradient. The shape information of the object is extracted using the edge detection algorithm. The Roberts filter is used for crack extraction. Two 2×2 convolution kernels are used, one for detecting horizontal edges and the other for detecting vertical edges. The convolution kernel is convolved with the image to calculate the horizontal and vertical gradient values ​​of each pixel point respectively. By calculating the gradients in the two directions, the gradient amplitude and direction of each pixel point are obtained, and the edge information in the image is depicted. After merging the horizontal and vertical gradients, the overall gradient image of the pixel point is obtained, as shown in formulas (1)-(3). The larger the value is, the larger the gradient value of the pixel is. By setting the threshold, all pixels with high gradient amplitude are selected to extract the edge contour of the crack in the image. (1) (2) (3) Among them: x, y represents the image The horizontal and vertical coordinates of the image The (x, y) position of the local window, P1-P9 represents the pixel value of the local window, and Representing images The horizontal gradient at the (x,y) position, and Representing images The vertical gradient at the (x,y) position, For images The first-order pixel gradient at position (x,y); 3.3. Supervised classification is to classify color images into different categories based on their features. It uses a supervised classification method based on the maximum likelihood method. This method assumes that the pixel features of each category follow a known probability distribution, including a Gaussian distribution. First, a training set is manually collected and divided into categories including cracks, vegetation, and rock mass. The collected training data set is analyzed to estimate the parameters of each category. The parameters include a mean vector and a covariance matrix. For each pixel to be classified, its probability density function value in each category is calculated based on its eigenvalue. The probability density function value represents the probability that the pixel belongs to a certain category. Subsequently, by comparing the probability density function values, the category with the largest probability is selected as the classification result of the pixel.

[0008] As an optimization: the step 4 is specifically as follows; 4.1. Morphological restoration. First, the input image is dilated. The dilation operation uses a structural element, i.e., a fixed-size rectangular frame, to perform a maximum value operation on the local area of ​​the image. The structural element is started from the upper left corner of the image and the entire image is scanned by sliding the structural element. For each pixel area covered by the structural element, the center of the structural element is compared with the pixels in the area. If there is any pixel point in the area that is a true value, i.e., a crack pixel point, the center of the structural element is assigned a true value. 4.2. After the dilation operation, the dilated image is subjected to an erosion operation. The erosion operation uses the structural element to perform a local minimum operation on the image. Similarly, the structural element is started from the upper left corner of the image and the entire image is scanned by sliding the structural element. For each pixel area covered by the structural element, the center of the structural element is compared with the pixels in the area. When all the pixels in the area are true values, the central pixel value of the structural element is assigned a true value, otherwise a false value is assigned. 4.3. Length filtering: Specifically, in the initially extracted crack image, according to Eps= , the parameter value of min_samples=2 is used for pixel clustering to form multiple crack units. After DBSCAN clustering, the pixels in the crack image will be divided into different clusters, each cluster represents a crack unit. Then, each crack unit is traversed to calculate the diagonal length of the minimum circumscribed rectangle. At the same time, a threshold is set to screen the crack units. The crack units whose diagonal length of the minimum circumscribed rectangle is less than the set threshold will be identified as noise or invalid cracks and removed; 4.4 Directional filtering, specifically: for each pixel in the binary image that is judged to be a crack, take the pixel as the circle point, sample the circle with a set radius, search for nearby pixels, and calculate the covariance matrix and eigenvalue of the pixel point , the eigenvalue represents the degree of change of the data in different directions, so the eigenvalue ratio is used as the directional index F of the pixel point, as shown in formula (4). At the same time, the directional threshold is set according to the image crack characteristics, and the pixels that do not meet the requirements are set as the background; 𝐹 = (4)

[0009] As an optimization: the step five is specifically as follows; The steps of fusing multi-dimensional crack information with the minimum risk Bayesian probability are as follows: The steps of the Bayesian probability method for minimum risk are: 5.1. It is necessary to clearly define the number of features n of the input sample and the types of permutations and combinations to obtain each feature ; 5.2. Next, determine the output category, that is, the final Bayesian decision classification result ; 5.3. Determine the prior probability distribution of various decision-making results in different scenarios by statistically analyzing existing historical cases ; 5.4. Statistical distribution probability of different features under each decision ; 5.5. Use the Bayesian formula to calculate each feature combination Under different conditions, different classification results The posterior probability (5); in, Indicates when The sample characteristics at represents the total number of sample features, Indicates when is the classification result of Bayesian decision, m represents the total number of classification results of Bayesian decision, represents the distribution probability of different features under each decision, Prior probability distribution; 5.6. Determine the risk factor under different decision outcomes ,in Indicates that when the decision classifies feature x as class, and x belongs to The loss coefficient that may be caused by class time, and The value range is 1 to m; 5.7. Calculate the classification risk of each class according to the formula , select the classification result with the smallest risk as The final decision (6) Among them, m represents the total number of Bayesian decision classification results, Indicates that when the decision classifies feature x as class, and x belongs to The loss coefficient that may be caused by class time, Indicates that the Y feature is classified as The posterior probability of the class.

[0010] The beneficial effects of the present invention are: 1. The three digital image-based crack automatic identification algorithms proposed in this method have shown significant applicability and effectiveness in large-scale, multi-morphological crack identification and investigation. They also show high robustness when processing crack images of different scales, shapes and background conditions, and their accuracy can stably reach more than 90%. 2. The filtering algorithm proposed in this method can effectively reduce the background noise of the cracks and improve the overall accuracy of the model. Direction and length filtering can effectively remove complex noise while minimizing the false damage of real cracks. The accuracy of each model is improved by 7.5%-55.3%, making the crack identification results more realistic and reliable. 3. The minimum risk Bayesian probability model fusion method proposed in this method can significantly improve the crack identification accuracy of a single model. The accuracy of the fused model can reach 95%. It can also flexibly adjust the weights and risk factors of different models according to actual conditions to achieve better identification effects. 4. This method combines the crack identification algorithm with drone inspection technology, which greatly improves the degree of automation and significantly reduces labor costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 It is a principle block diagram of the present invention.

[0012] Figure 2 It is a schematic diagram of the grayscale redistribution result in the present invention.

[0013] Figure 3 Schematic diagram of the threshold segmentation principle in the present invention, a- is the crack grayscale image; b- is the crack and background grayscale features; c- is the crack image after threshold segmentation.

[0014] Figure 4 It is a schematic diagram of the edge detection principle in the present invention.

[0015] Figure 5 This is a schematic diagram of the comparison of morphology before and after restoration in the present invention.

[0016] Figure 6 It is a schematic diagram for comparing the length before and after filtering in the present invention.

[0017] Figure 7 It is a schematic diagram of the comparison before and after directional filtering in the present invention. DETAILED DESCRIPTION

[0018] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby making a clearer and more definite definition of the protection scope of the present invention.

[0019] like Figure 1 As shown: A reservoir slope crack identification method based on drone inspection and digital images, the specific steps are: 1. Data acquisition Different types of hangars have different functional focuses, so during inspection work, comprehensive consideration should be given to selecting the most suitable drone hangar for the current inspection task. According to the on-site environmental conditions of the reservoir dam area and the inspection technical requirements, and taking into account the cost and flexibility of deployment, installation and flight, it is planned to use an automatic hangar adapted to the multi-rotor drone platform. The final selection is as follows: The drone automatic hangar uses the DJI Airport, which has a highly integrated fuselage design, integrating ultra-wide-angle surveillance cameras, anemometers, rain gauges, communication antennas, RTK modules, UPS power supplies, etc. The DJI Airport has strong environmental adaptability and can persist even in unmanned areas with harsh climates. It has built-in lightning protection function, and the entire airport reaches IP55 (refer to IEC 60529 standard) protection level, and the longest maintenance interval can be up to 6 months.

[0020] Table 1 DJI Airport Performance Parameters The drones used are the Matrice M30 series, which are terminal devices in DJI's airport system. They have two versions, M30 and M30T. The difference is that the M30T has a thermal imaging payload and can perform thermal infrared inspections at night. The airport version of the M30 series drone is equipped with a charging module that can be docked with the airport, achieving perfect integration with the airport. It can be quickly charged without removing the battery, greatly improving the reliability of airport operations. At the same time, the TB30 smart battery it carries has a cycle count of more than 400 times, providing capability guarantee for the execution of a large number of frequent on-site tasks.

[0021] Table 2 Performance parameters of the M30T UAV Different from the "strip-shaped" operation range of power inspection and pipeline inspection, the inspection task area of ​​crack identification work is mainly "surface-shaped", so the drone inspection task is carried out through zoning planning. After the zoning planning is completed, the mission route is planned through the data management platform. In the data management platform, the heading overlap rate, lateral overlap rate, flight altitude, ground resolution and other elements of the route can be designed according to different mission requirements, and multiple routes can be planned according to different inspection frequencies and then uploaded. Finally, the inspection task can be automatically executed at the time set according to the requirements. After the drone completes the inspection task and returns to the hangar, the photos collected for this task will be automatically uploaded to the cloud system, and the original image data can be downloaded remotely for data processing.

[0022] 2. Data preprocessing At present, most researchers of image-based crack extraction use color (RGB value) as the main feature to distinguish cracks from background. Since most of the reservoir slope areas belong to high mountain canyon areas, the vegetation coverage rate is high and the working environment is complex, the optical image data collected in this area has the characteristics of large background noise interference and significant interference from climate and light. In order to minimize the influence of noise generated by these factors, the grayscale-grayscale enhancement-bilateral filtering processing method is first used to pre-process the data.

[0023] Firstly, according to the digital television signal standard, the original RGB image is averaged with a weight of red: green: blue = 0.298: 0.587: 0.114, and a single-channel grayscale image is generated from the three-channel RGB image to eliminate the additional complexity that may be introduced by color information and highlight the texture, shape and structure information of the image.

[0024] Secondly, the histogram equalization method is used to enhance the grayscale of the image. First, a matrix of a certain size is used to collect the entire image to understand the pixel distribution of each grayscale level. Finally, grayscale mapping is performed according to the cumulative distribution function to assign each pixel value in the original image to a new grayscale level, thereby achieving grayscale enhancement. The grayscale redistribution result is shown in the figure below. Figure 2 This method can effectively enhance the visual performance of the image, improve the quality of the image, highlight the details in the image, and make the image more suitable for subsequent image processing and analysis tasks.

[0025] Finally, the image is subjected to bilateral filtering for denoising. The advantage of bilateral filtering is that it can effectively handle different types of noise, including Gaussian noise, salt and pepper noise, etc., so that the image will not be blurred or distorted while denoising. By properly selecting the size and standard deviation of the filter, the details and edge information of the image can be retained as much as possible while removing noise.

[0026] 3. Crack identification based on digital images A digital image is an image composed of discrete pixels, each of which contains specific color or grayscale information. These pixels form the overall visual appearance of the image in a certain arrangement. The more common digital images are color images, binary images, and grayscale images. Each pixel of a color image is usually composed of three color channels, red, green, and blue, according to different values. A binary image is an image type that contains only two colors, black and white. Each pixel can only take two values, usually 0 and 1, representing black and white. A grayscale image means that the grayscale value of each pixel represents the brightness information in the image, usually ranging from 0 (black) to 255 (white). Digital images are also stored and processed in pixels in computers. Therefore, depending on the pixel unit, the method of identifying cracks is also different, which can be basically divided into three categories: 1. Threshold segmentation; 2. Edge detection; 3. Supervised classification.

[0027] Threshold segmentation is the most commonly used, simplest and most effective method in crack identification models

[17] . Because cracks are sunken on the ground, they usually produce black crack shadows at the bottom of the cracks when illuminated by oblique sunlight, making them often appear as black pixels in grayscale images. After the digital images collected by drones are grayscaled, a certain grayscale threshold is set, and grayscale values ​​less than H are set as foreground, and grayscale values ​​greater than H are set as background, thereby achieving the effect of crack extraction, such as Figure 3 shown.

[0028] Edge detection is the process of judging whether a pixel belongs to the edge of an object by the change in grayscale gradient. The shape information of the object can be effectively extracted by using edge detection algorithms. Different edge detection algorithms actually use different edge filters. Common filters include Sobel, Roberts, Robinson and Frei filters. [18-20] , these filters extract the edge contour of the region by finding the part of the image where the gradient of the pixel gray value changes significantly and selecting it. This study uses the Roberts filter to extract cracks. It uses two 2×2 convolution kernels, one for detecting horizontal edges and the other for detecting vertical edges. These convolution kernels are convolved with the image to calculate the horizontal and vertical gradient values ​​of each pixel. By calculating the gradients in these two directions, the gradient amplitude and direction of each pixel can be obtained, thereby depicting the edge information in the image. After merging the horizontal and vertical gradients, the overall gradient image of the pixel can be obtained, as shown in formulas (1)-(3). The larger the value is, the larger the gradient value of the pixel is. By setting the threshold, all pixels with higher gradient amplitudes are selected to extract the edge contour of the crack in the image. Figure 4 shown.

[0029] (1) (2) (3) Among them: P1-P9 represents the pixel value of the local window, Represents the horizontal gradient change, Represents the vertical gradient change, is the first-order pixel gradient.

[0030] Supervised classification is to classify color images into different categories based on their features. This method uses the maximum likelihood supervised classification method, which assumes that the pixel features of each category follow a known probability distribution, usually a Gaussian distribution. First, the training set is manually collected and divided into cracks, vegetation, rock mass or other different categories, such as Figure 7 As shown, the collected training data set is analyzed to estimate the parameters of each category, including the mean vector and the covariance matrix. For each pixel to be classified, the probability density function value under each category is calculated according to its eigenvalue. This value represents the probability that the pixel belongs to a certain category. Subsequently, by comparing these probability density function values, the category with the largest probability is selected as the classification result of the pixel.

[0031] 4. Crack pixel filtering There will still be some noise points in the crack image after the initial extraction, which will not only affect the subsequent analysis, but also affect the accurate assessment of the real cracks. Therefore, in order to improve the accuracy and reliability of crack detection, the extracted crack image must be filtered. Filtering technology can effectively suppress noise and retain key crack information, thereby providing a clearer and more accurate crack image. This method proposes two filtering methods for the binary crack image after initial extraction based on the crack morphology of the reservoir slope, namely length filtering and direction filtering.

[0032] 4.1 Morphological Restoration The crack pixels obtained by the edge detection model can only represent the edge features of the cracks, among which a large number of pixels belonging to the cracks are judged as the background, and the other two models may also have a small number of real crack pixels that are misjudged as the background. Therefore, in order to repair the real crack pixels, the morphological closing operation method is used for the initially identified crack binary image. This method is mainly used to fill small holes or link interrupted places. The main steps of the morphological closing operation are divided into two steps: expansion and corrosion: 1. First, dilate the input image. The dilation operation uses a structural element, i.e. a fixed-size rectangular box, to perform a maximum value operation on a local area of ​​the image. The structural element is started from the upper left corner of the image and the entire image is scanned by sliding the structural element. For each pixel area covered by the structural element, the center of the structural element is compared with the pixels in the area. If any pixel in the area is a true value (i.e. a crack pixel), the center of the structural element is assigned a true value.

[0033] 2. After the dilation operation, the dilated image is eroded. The erosion operation uses the structural element to perform the minimum value operation on the local area of ​​the image. Similarly, the structural element starts from the upper left corner of the image and scans the entire image by sliding the structural element. For each pixel area covered by the structural element, the center of the structural element is compared with the pixels in the area. Only when all the pixels in the area are true values, the central pixel value of the structural element is assigned a true value, otherwise it is assigned a false value. The crack repair effect is as follows Figure 5 shown.

[0034] 4.2 Length Filtering In the initially extracted crack image, there are often some isolated, small noise points. In order to effectively process these noise points, according to Eps= , the parameter value of min_samples=2 is used for pixel clustering. The purpose of this parameter value is to cluster adjacent pixels with at least one common endpoint in order to form multiple crack units. After DBSCAN clustering, the pixels in the crack image will be divided into different clusters, each cluster representing a crack unit. Next, each crack unit is traversed to calculate the diagonal length of its minimum circumscribed rectangle. At the same time, a threshold is set to filter the crack units. Crack units whose diagonal length of the minimum circumscribed rectangle is less than the set threshold will be identified as noise points or invalid cracks and will be removed. The DBSCAN algorithm can effectively divide data points into core points, boundary points, and noise points. Clustering can better distinguish noise points from valid signals. This helps to more accurately identify and process noise points in subsequent length filtering. The effect of length filtering is as follows Figure 6 shown.

[0035] 4.3 Directional Filtering Cracks usually appear in images in the form of elongated shapes with obvious directional characteristics, while background noise is often distributed in irregular shapes. Therefore, for each pixel in the binary image that is judged to be a crack, the pixel is used as a circle point, and a circle with a certain radius is sampled to search for nearby pixels, and the covariance matrix and eigenvalues ​​of these pixels are calculated. , the eigenvalue represents the degree of change of the data in different directions, so the eigenvalue ratio is used as the directional index F of the pixel point, as shown in Formula 4. At the same time, the directional threshold is set according to the image crack characteristics, and the pixels that do not meet the requirements are set as the background. The directional filtering effect is shown in Figure 7 shown.

[0036] 𝐹 = (4) 5. Model Fusion In order to improve the accuracy and robustness of crack identification, this method adopts the minimum risk Bayesian probability method to organically integrate the three models of threshold segmentation, edge detection and supervised classification. This method overcomes the limitations of a single crack identification model and gives full play to the advantages of each model, thereby improving the performance and robustness of the crack identification system.

[0037] Bayesian decision theory is a decision-making method based on Bayesian probability theory, which is used to make the best decision under uncertainty. The theory is based on the Bayesian formula, combines the information of prior probability and observation data, calculates the posterior probability, and selects the decision result with the maximum posterior probability. However, this method does not take into account the risks brought by each decision. To this end, the risk coefficient of each decision needs to be included in the final selection to achieve the fusion of crack information under different risk conditions. The steps of the minimum risk Bayesian probability fusion of multidimensional crack information proposed by this method are as follows: 5.1. First, we need to clarify the number of features (n) of the input sample and their permutations and combinations, so as to obtain each feature 5.2. Next, determine the output category, which is the final Bayesian decision classification result 5.3. Determine the prior probability distribution of various decision-making results in different scenarios by statistically analyzing existing historical cases .

[0038] 5.4. Statistical distribution probability of different features under each decision .

[0039] 5.5. Use the Bayesian formula to calculate each feature combination Under different conditions, different classification results The posterior probability .

[0040] (5) 5.6. Determine the risk factor under different decision outcomes ,in Indicates that when the decision classifies feature x as class, but in fact x belongs to The loss coefficient that may be caused by class time, and The value range is 1 to m.

[0041] 5.7. Finally, the classification risk of each class is calculated according to the formula , select the classification result with the smallest risk as final decision.

[0042] (6) The method of this embodiment shows obvious advantages in the field of automatic crack identification. By combining multiple algorithm models, background noise processing and model fusion strategies, the accuracy and stability of reservoir slope crack identification are successfully improved. At the same time, the combination with drone inspection technology greatly improves the degree of automation and significantly reduces labor costs. First, after obtaining the orthophoto map through drone inspection, three algorithm models using image pixel gradient, grayscale value and RGB value as indicators are used to achieve preliminary identification, overcoming the problem of low accuracy of a single algorithm model under complex backgrounds. Secondly, by processing background noise through morphological repair and filtering algorithms, interference is effectively reduced, and the accuracy and reliability of crack identification are improved. Most importantly, the Bayesian probability model fusion method with minimum risk is introduced to fuse the recognition results of different models at the decision level, and finally the accuracy of reservoir slope crack identification can be increased to 95% or more. The application of this comprehensive technical solution brings new practical solutions to the field of automatic crack identification, with broad application prospects and technical effects, and is expected to promote further development and application in this field.

Claims

1. A reservoir slope crack identification method based on drone inspection and digital images, characterized by: The specific steps are: Step 1: Data acquisition, using drones to collect reservoir slope image data; Step 2: Image preprocessing, using grayscale, grayscale enhancement and bilateral filtering methods to preprocess the data; Step 3: Crack identification: using threshold segmentation, edge detection and supervised classification methods to identify the cracks on the reservoir slope in the image; Step 4: Crack pixel filtering, using morphological closing operation, length filtering and direction filtering to filter the extracted crack image; Step 5: Model fusion, using the minimum risk Bayesian probability method to fuse the three models of threshold segmentation, edge detection and supervised classification.

2. The reservoir slope crack identification method based on drone inspection and digital images according to claim 1 is characterized by: The step 2 specifically comprises: 2.

1. According to the digital television signal standard, the original RGB image is averaged with a weight of red: green: blue = 0.298: 0.587: 0.114, and a single-channel grayscale image is generated from the three-channel RGB image to eliminate the additional complexity introduced by color information and highlight the texture, shape and structure information of the image; 2.

2. Use the histogram equalization method to enhance the grayscale of the image. First, use the matrix to collect the entire image and identify the pixel distribution of each grayscale level. Finally, perform grayscale mapping based on the cumulative distribution function to assign each pixel value in the original image to a new grayscale level to achieve grayscale enhancement. 2.

3. Perform bilateral filtering to denoise the image, including Gaussian noise and salt and pepper noise, so that the image will not be blurred or distorted while denoising. By selecting the size and standard deviation of the filter, the details and edge information of the image can be retained while removing noise.

3. The reservoir slope crack identification method based on drone inspection and digital images according to claim 1 is characterized by: The step three is specifically as follows: 3.

1. Threshold segmentation: Specifically, after the digital image collected by the drone is grayed, the grayscale threshold H is set, and the grayscale value less than H is set as the foreground, and the grayscale value greater than H is set as the background, so as to realize crack extraction; 3.

2. Edge detection is a process of judging whether a pixel belongs to the edge of an object by the change in grayscale gradient. The shape information of the object is extracted using the edge detection algorithm. The Roberts filter is used for crack extraction. Two 2×2 convolution kernels are used, one for detecting horizontal edges and the other for detecting vertical edges. The convolution kernel is convolved with the image to calculate the horizontal and vertical gradient values ​​of each pixel point respectively. By calculating the gradients in the two directions, the gradient amplitude and direction of each pixel point are obtained, and the edge information in the image is depicted. After merging the horizontal and vertical gradients, the overall gradient image of the pixel point is obtained, as shown in formulas (1)-(3). The larger the value is, the larger the gradient value of the pixel is. By setting the threshold, all pixels with high gradient amplitude are selected to extract the edge contour of the crack in the image. (1) (2) (3) Among them: x, y represents the image The horizontal and vertical coordinates of the image The (x, y) position of the local window, P1-P9 represents the pixel value of the local window, and Representing images The horizontal gradient at the (x,y) position, and Representing images The vertical gradient at the (x,y) position, For images The first-order pixel gradient at position (x,y); 3.

3. Supervised classification is to classify color images into different categories based on their features. It uses a supervised classification method based on the maximum likelihood method. This method assumes that the pixel features of each category follow a known probability distribution, including a Gaussian distribution. First, a training set is manually collected and divided into categories including cracks, vegetation, and rock mass. The collected training data set is analyzed to estimate the parameters of each category. The parameters include a mean vector and a covariance matrix. For each pixel to be classified, its probability density function value in each category is calculated based on its eigenvalue. The probability density function value represents the probability that the pixel belongs to a certain category. Subsequently, by comparing the probability density function values, the category with the largest probability is selected as the classification result of the pixel.

4. The reservoir slope crack identification method based on drone inspection and digital images according to claim 1 is characterized by: The step 4 is specifically as follows: 4.

1. Morphological restoration. First, the input image is dilated. The dilation operation uses a structural element, i.e., a fixed-size rectangular frame, to perform a maximum value operation on the local area of ​​the image. The structural element is started from the upper left corner of the image and the entire image is scanned by sliding the structural element. For each pixel area covered by the structural element, the center of the structural element is compared with the pixels in the area. If there is any pixel point in the area that is a true value, i.e., a crack pixel point, the center of the structural element is assigned a true value. 4.

2. After the dilation operation, the dilated image is subjected to an erosion operation. The erosion operation uses the structural element to perform a local minimum operation on the image. Similarly, the structural element is started from the upper left corner of the image and the entire image is scanned by sliding the structural element. For each pixel area covered by the structural element, the center of the structural element is compared with the pixels in the area. When all the pixels in the area are true values, the central pixel value of the structural element is assigned a true value, otherwise a false value is assigned. 4.

3. Length filtering: Specifically, in the initially extracted crack image, according to Eps= , the parameter value of min_samples=2 is used for pixel clustering to form multiple crack units. After DBSCAN clustering, the pixels in the crack image will be divided into different clusters, each cluster represents a crack unit. Then, each crack unit is traversed to calculate the diagonal length of the minimum circumscribed rectangle. At the same time, a threshold is set to screen the crack units. The crack units whose diagonal length of the minimum circumscribed rectangle is less than the set threshold will be identified as noise or invalid cracks and removed; 4.4 Directional filtering, specifically: for each pixel in the binary image that is judged to be a crack, take the pixel as the circle point, sample the circle with a set radius, search for nearby pixels, and calculate the covariance matrix and eigenvalue of the pixel point , the eigenvalue represents the degree of change of the data in different directions, so the eigenvalue ratio is used as the directional index F of the pixel point, as shown in formula (4). At the same time, the directional threshold is set according to the image crack characteristics, and the pixels that do not meet the requirements are set as the background; 𝐹= (4)。 5. The reservoir slope crack identification method based on drone inspection and digital images according to claim 1 is characterized by: The step five is specifically as follows: The steps of the Bayesian probability method for minimum risk are: 5.

1. It is necessary to clearly define the number of features n of the input sample and the types of permutations and combinations to obtain each feature ; 5.

2. Next, determine the output category, that is, the final Bayesian decision classification result ; 5.

3. Determine the prior probability distribution of various decision-making results in different scenarios by statistically analyzing existing historical cases ; 5.

4. Statistical distribution probability of different features under each decision ; 5.

5. Use the Bayesian formula to calculate each feature combination Under different conditions, different classification results The posterior probability (5)； in, Indicates when The sample characteristics at represents the total number of sample features, Indicates when is the classification result of Bayesian decision, m represents the total number of classification results of Bayesian decision, represents the distribution probability of different features under each decision, Prior probability distribution; 5.

6. Determine the risk factor under different decision outcomes ,in Indicates that when the decision classifies feature x as class, and x belongs to The loss coefficient that may be caused by the class time, and The value range is 1 to m; 5.

7. Calculate the classification risk of each class according to the formula , select the classification result with the smallest risk as The final decision (6) Among them, m represents the total number of Bayesian decision classification results, Indicates that when the decision classifies feature x as class, and x belongs to The loss coefficient that may be caused by the class time, Indicates that the Y feature is classified as The posterior probability of the class.

Images