Dam crack detection system based on unmanned aerial vehicle technology

Abstract

The invention belongs to the field of dam crack detection, and discloses a dam crack detection system based on an unmanned aerial vehicle technology, which comprises an unmanned aerial vehicle and an image processing terminal, the unmanned aerial vehicle is used for shooting the slope of the dam to obtain a plurality of slope images; the image processing terminal is used for identifying the slope image to obtain a detection result; the shooting of the slope of the dam comprises the following steps: obtaining a plurality of shooting heights based on the height of the slope of the dam; the unmanned aerial vehicle horizontally moves at each shooting height, the distance between the unmanned aerial vehicle and the slope of the dam in the horizontal direction is kept the same in the moving process, and the slope of the dam is shot once every fixed moving distance to obtain a slope image. According to the invention, the efficiency of detecting the crack of the slope of the dam is greatly improved.

Description

Technical Field

[0001] The present invention relates to the field of dam crack detection, and in particular to a dam crack detection system based on unmanned aerial vehicle technology. Background Art

[0002] Dam crack detection is the process of monitoring and evaluating dam structures to identify and analyze cracks and other potential defects. This process is critical to ensuring the safety and long-term stability of dams, especially when they are subjected to water pressure and other environmental factors.

[0003] The traditional method for detecting cracks on the slope of a dam is mainly visual detection, which has the problem of low detection efficiency and needs to be improved. Summary of the invention

[0004] The purpose of the present invention is to disclose a dam crack detection system based on drone technology to solve the technical problems raised in the background technology.

[0005] In order to achieve the above object, the present invention provides the following technical solutions:

[0006] The present invention provides a dam crack detection system based on UAV technology, including a UAV and an image processing terminal;

[0007] The drone is used to photograph the slope of the dam and obtain multiple slope images;

[0008] The image processing terminal is used to identify the inclined surface image and obtain the detection result;

[0009] Among them, the slope of the dam was photographed, including:

[0010] Acquire multiple shooting heights based on the height of the slope of the dam;

[0011] The drone moves horizontally at each shooting height. During the movement, it maintains the same horizontal distance from the slope of the dam. Every fixed moving distance, it takes a picture of the slope of the dam to obtain an image of the slope.

[0012] Preferably, the image processing terminal is a computer, which can establish a connection with the drone via WiFi or a data cable. After the connection is established, the computer can access the captured slope images stored in the storage device of the drone.

[0013] Preferably, a plurality of shooting heights are obtained based on the height of the slope of the dam, including:

[0014] Let H represent the height of the slope of the dam;

[0015] D represents the horizontal distance between the drone and the slope of the dam when taking the photo;

[0016] N represents the number of shooting heights, and N enables the slope images obtained at adjacent shooting heights to overlap in the vertical direction when the UAV shoots the slope of the dam based on D;

[0017] The first shooting height is

[0018] The nth shooting height is

[0019] Preferably, the relationship between the area of ​​the overlapping region and the area of ​​the oblique surface image is:

[0020]

[0021] Sa represents the area of ​​the overlapping region in the oblique surface image, Sb represents the area of ​​the oblique surface image, and h represents the set ratio.

[0022] Preferably, the value of h is 0.5.

[0023] Preferably, identifying the inclined surface image to obtain the detection result includes:

[0024] Preprocessing each inclined plane image respectively to obtain a preprocessed inclined plane image;

[0025] Stitching the preprocessed oblique images to obtain a stitched image;

[0026] Use a deep learning-based detection algorithm to identify the spliced ​​image and obtain the detection result.

[0027] Preferably, the inclined surface image is preprocessed, including:

[0028] Performing distortion correction processing on the oblique plane image to obtain a distortion-corrected image;

[0029] Performing illumination correction processing on the distortion-corrected image to obtain an illumination-corrected image;

[0030] The illumination correction image is subjected to denoising to obtain a preprocessed oblique image.

[0031] Preferably, the pre-processed oblique images are stitched together to obtain a stitched image, comprising:

[0032] Use the following algorithm to stitch images and obtain a stitched image:

[0033] Perform feature point detection on the two images p1 and p2 to be stitched, and obtain the feature points in p1 and p2;

[0034] Use the matching algorithm to obtain the same feature points in p1 and p2 to obtain multiple feature point pairs;

[0035] Eliminate erroneous feature point pairs from multiple feature point pairs to obtain feature point pairs for image registration;

[0036] Estimate the geometric transformation matrix between p1 and p2 based on the feature point pairs used for image registration;

[0037] Based on the geometric transformation matrix, p1 is transformed to the coordinate system where p2 is located to obtain image p3. In the coordinate system where p2 is located, the grayscale values ​​of the pixels in p2 and p3 are weighted averaged to obtain the grayscale values ​​of the pixels in the spliced ​​image.

[0038] Preferably, the matching algorithm includes any one of FLANN and KNN.

[0039] Preferably, removing erroneous feature point pairs from a plurality of feature point pairs to obtain feature point pairs for image registration includes:

[0040] The RANSAC algorithm is used to eliminate erroneous feature point pairs from multiple feature point pairs to obtain feature point pairs for image registration.

[0041] Beneficial effects:

[0042] Compared with the existing crack detection method, the present invention obtains multiple shooting heights according to the height of the dam, and then uses a drone to move horizontally at each shooting height to shoot the slope of the dam, thereby acquiring the slope image of the dam. After that, a detection algorithm based on deep learning is used to identify the spliced ​​image obtained by splicing the slope images, thereby greatly improving the efficiency of detecting cracks on the slope of the dam. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for describing the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0044] Figure 1 This is a schematic diagram of a dam crack detection system based on drone technology according to the present invention.

[0045] Figure 2 Schematic diagram of the process of preprocessing the inclined surface image. DETAILED DESCRIPTION

[0046] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0047] like Figure 1 As shown, the present invention provides a dam crack detection system based on UAV technology, including a UAV and an image processing terminal;

[0048] The drone is used to photograph the slope of the dam and obtain multiple slope images;

[0049] The image processing terminal is used to identify the inclined surface image and obtain the detection result;

[0050] Among them, the slope of the dam was photographed, including:

[0051] Acquire multiple shooting heights based on the height of the slope of the dam;

[0052] The drone moves horizontally at each shooting height. During the movement, it maintains the same horizontal distance from the slope of the dam. Every fixed moving distance, it takes a picture of the slope of the dam to obtain an image of the slope.

[0053] The present invention obtains multiple shooting heights according to the height of the dam, and then uses a drone to move in a horizontal direction at each shooting height to shoot the slope of the dam, thereby acquiring the slope image of the dam. Afterwards, a detection algorithm based on deep learning is used to identify the spliced ​​image obtained by splicing the slope images, thereby greatly improving the efficiency of detecting cracks on the slope of the dam.

[0054] Preferably, the image processing terminal is a computer, which can establish a connection with the drone via WiFi or a data cable. After the connection is established, the computer can access the captured slope images stored in the storage device of the drone.

[0055] Specifically, the drone can be controlled by the staff responsible for inspection, or it can fly automatically after the route is planned.

[0056] Preferably, a plurality of shooting heights are obtained based on the height of the slope of the dam, including:

[0057] Let H represent the height of the slope of the dam;

[0058] D represents the horizontal distance between the drone and the slope of the dam when taking the photo;

[0059] N represents the number of shooting heights, and N enables the slope images obtained at adjacent shooting heights to overlap in the vertical direction when the UAV shoots the slope of the dam based on D;

[0060] The first shooting height is

[0061] The nth shooting height is

[0062] Specifically, D can be obtained as follows:

[0063] Get the horizontal plane Q where the drone is located;

[0064] Obtain the curve S formed by the intersection of the horizontal plane Q and the slope of the dam;

[0065] Connect the point node of the curve S and the drone to obtain a line segment Snode, which can be perpendicular to the moving direction of the drone; the length of Snode is D.

[0066] Preferably, the fixed moving distance can ensure that two oblique images obtained at the same shooting height and with adjacent shooting positions have overlapping areas in the horizontal direction.

[0067] Preferably, the relationship between the area of ​​the overlapping region and the area of ​​the oblique surface image is:

[0068]

[0069] Sa represents the area of ​​the overlapping region in the oblique surface image, Sb represents the area of ​​the oblique surface image, and h represents the set ratio.

[0070] Preferably, the value of h is 0.5.

[0071] Preferably, identifying the inclined surface image to obtain the detection result includes:

[0072] Preprocessing each inclined plane image respectively to obtain a preprocessed inclined plane image;

[0073] Stitching the preprocessed oblique images to obtain a stitched image;

[0074] Use a deep learning-based detection algorithm to identify the spliced ​​image and obtain the detection result.

[0075] Specifically, the detection algorithm based on deep learning may be the YOLOv8 algorithm.

[0076] The training process of the YOLOv8 algorithm is as follows:

[0077] 1. Prepare the dataset:

[0078] Data Collection:

[0079] To use YOLOv8 to detect cracks in the slope of the dam, you first need to collect a large number of labeled images to ensure that the images contain different types of cracks, different angles, different lighting conditions, different sizes, etc.

[0080] Image source: Images of the dam surface taken by drones or on-site monitoring pictures of the dam can be used.

[0081] Image quality requirements: The image should be as clear as possible, including the characteristics of the cracks, and ensure that the cracks are sufficiently prominent in the image.

[0082] Data diversity: In order for the model to have better generalization ability, the dataset should contain a variety of backgrounds, lighting, weather conditions, images taken from different angles, etc.

[0083] Data annotation:

[0084] Use annotation tools to select cracks in the image and annotate them. Commonly used annotation tools include LabelImg, LabelMe, etc.

[0085] Annotation format: YOLO requires the annotation file to be stored in .txt format. Each image corresponds to a text file, and each line in the file represents the category and coordinates of a crack box.

[0086] For each target (crack), the annotation format of YOLOv8 is:

[0087] <class_id><center_x><center_y> <width> <height>

[0088] class_id: The ID of the crack category. If there is only one category of cracks, the ID is 0.

[0089] center_x and center_y: The center point coordinates of the crack bounding box, normalized to the range [0,1].

[0090] width and height: The width and height of the crack bounding box, also normalized to the range [0,1].

[0091] For example, annotate the location of a crack in an image as follows: 0 0.5 0.5 0.2 0.1

[0093] Represents a crack with ID 0, whose center point coordinates are the center of the image, width is 0.2, and height is 0.1.

[0094] Data partitioning:

[0095] Divide the dataset into training set, validation set and test set, usually in the following proportions:

[0096] Training set: 70-80% (for model training);

[0097] Validation set: 10-15% (for model tuning);

[0098] Test set: 10-15% (used to evaluate model performance);

[0099] 2. Install the YOLOv8 environment:

[0100] Installation dependencies:

[0101] YOLOv8 can be trained and inferred through the official library provided by Ultralytics. First, you need to install the relevant Python packages and dependencies:

[0102] pip install ultralytics;

[0103] 3. Configure training parameters

[0104] Data Configuration

[0105] Before starting training, you need to set up the configuration files required by YOLOv8. Create a .yaml file to define the structure of the dataset. This configuration file contains class labels and data paths. Assume there is a configuration file named crack_detection.yaml with the following content:

[0106] yamltrain: / path / to / train / images;

[0107] val: / path / to / val / images;

[0108] nc:1#Number of categories, crack has only one category

[0109] names:['crack']#Category name

[0110] train: Path to training images.

[0111] val: Path to the validation image.

[0112] nc: number of categories, here set to 1, because the present invention only identifies cracks.

[0113] names: Category names. There is only one category "crack".

[0114] Configuration file (yaml) storage location:

[0115] The crack_detection.yaml configuration file needs to be placed in the training directory, and the path can be set according to the actual situation.

[0116] 4. Select a pre-trained model:

[0117] YOLOv8 provides multiple pre-trained models, and you can choose a suitable model as the basis for transfer learning. You can choose to start training from the model provided in the YOLOv8 GitHub official library. The choice of pre-trained model will be determined based on factors such as the complexity of the training data and hardware resources.

[0118] Pre-trained model download:

[0119] YOLOv8 provides pre-trained models of different sizes (such as YOLOv8n, YOLOv8s, YOLOv8m, YOLOv8l, YOLOv8x, etc.), and you can choose the appropriate pre-trained model according to your needs.

[0120] 5. Start training the model:

[0121] Start the YOLOv8 training process with the following command:

[0122] yolo task=detect mode=train model=yolov8n.pt data=crack_detection.yaml epochs=50batch=16imgsz=640

[0123] Command parameter description:

[0124] task=detect: specifies the task as target detection.

[0125] mode=train: indicates training.

[0126] model=yolov8n.pt: Select the pre-trained YOLOv8 model. yolov8n.pt is a smaller model. You can also choose a larger model, such as yolov8l.pt, based on the hardware performance and accuracy requirements.

[0127] data=crack_detection.yaml: specifies the dataset configuration file.

[0128] epochs=50: The number of training rounds (can be adjusted according to the size of the dataset and training results).

[0129] batch=16: batch size for each iteration (adjusted according to video memory).

[0130] imgsz=640: The size of the input image.

[0131] 6. Monitor the training process:

[0132] During the training process, YOLOv8 will output the loss function, precision, recall rate and other indicators for each epoch. These indicators can be used to judge the training progress of the model. If the accuracy is found to be reduced or the model is overfitting, the training process can be further optimized by adjusting the learning rate, increasing the data set, enhancing the data, etc.

[0133] 7. Model evaluation and tuning:

[0134] After training, you can use the validation set to evaluate the model's performance. If the model accuracy is not ideal, you can try the following methods to optimize it:

[0135] Increase the diversity of the dataset: Add more crack images, especially with different lighting conditions, viewing angles, crack morphologies, etc.

[0136] Data enhancement: Perform image enhancement, such as rotation, scaling, brightness adjustment, etc., to help the model generalize better.

[0137] Adjust hyperparameters: such as learning rate, batch size, training rounds, etc.

[0138] Fine-tuning: Fine-tune the model and train it for more rounds using a smaller learning rate.

[0139] 8. Reasoning and Testing

[0140] After training is complete, you can use the trained model for inference (prediction):

[0141] yolo task=detect mode=predict model=path / to / best_model.pt source=path / to / test / images

[0142] The source parameter can be the path to a test image or video file. After prediction, YOLOv8 will output an image or video with crack marks, which can be used for further analysis.

[0143] 9. Save and deploy

[0144] After the model training is completed and the test passes, you can save the model to the specified directory and deploy it according to your needs. Save the model:

[0145] yolo task=detect mode=export model=path / to / best_model.pt format=onnx

[0146] The model can be exported to ONNX format for easy deployment to other platforms.

[0147] Preferably, if Figure 2 , preprocess the inclined image, including:

[0148] Performing distortion correction processing on the oblique plane image to obtain a distortion-corrected image;

[0149] Performing illumination correction processing on the distortion-corrected image to obtain an illumination-corrected image;

[0150] The illumination correction image is subjected to denoising to obtain a preprocessed oblique image.

[0151] Specifically, when performing distortion correction processing on the oblique image, a perspective transformation correction method can be used for correction. By calculating the homography matrix between the original image and the target image, the perspective distortion of the image can be corrected.

[0152] Preferably, performing illumination correction processing on the distortion-corrected image to obtain the illumination-corrected image includes:

[0153] Adopting an adaptive partitioning algorithm to partition the distortion-corrected image to obtain multiple areas to be corrected;

[0154] Correct each area to be corrected respectively to obtain a corrected area;

[0155] All the correction areas are merged to obtain the illumination correction image.

[0156] The present invention performs illumination correction in a region-by-region manner, which can effectively reduce the probability of occurrence of events that lead to enhanced noise in local areas when performing illumination correction. This is because when illumination correction is performed directly in the distortion-corrected image, the noise of the pixel points in the corrected local area is easily enhanced, thereby affecting the accuracy of subsequent recognition.

[0157] Preferably, an adaptive partitioning algorithm is used to partition the distortion-corrected image to obtain a plurality of areas to be corrected, including:

[0158] The first step is to divide the distortion-corrected image into Z local regions of equal area, and store the obtained local regions into a set ZU;

[0159] The second step is to determine whether each local area in ZU needs to be further divided;

[0160] The third step is to store the local areas that need to be further divided into the set SU; and store the local areas that do not need to be further divided into the set FU;

[0161] The fourth step is to determine whether the number of local regions in SU is greater than or equal to 1. If so, proceed to the fifth step. If not, take the local region in FU as the region to be corrected.

[0162] Step 5: reset ZU to an empty set;

[0163] The sixth step is to divide each element in SU into Z local regions of equal area, store the obtained local regions into the set ZU, and proceed to the second step.

[0164] The present invention does not directly divide the distortion-corrected image into a plurality of to-be-corrected regions of the same area, because the size of the to-be-corrected region is difficult to determine. If the area is too small, it is easy to divide the region with lower contrast into too many to-be-corrected regions, which increases the probability of erroneously introducing edge information after correction. If the area is too large, it is easy to divide a plurality of regions with large contrast differences into the same correction region, which suppresses the original edge information after the partitioned illumination correction.

[0165] Therefore, the present invention realizes the adaptive determination of the area of ​​the area to be corrected by judging whether each local area needs to be further divided, so that after the illumination correction, more edge information can be retained and the introduction of erroneous edge information can be suppressed, which is conducive to obtaining a higher quality illumination correction image.

[0166] Preferably, determining whether each local area in the ZU needs to be further divided includes:

[0167] Use zu to represent the local area in ZU and calculate the judgment value of zu:

[0168]

[0169] Value zu Indicates the judgment value of zu, nzu indicates the number of pixels in zu, gb i represents the gray value of pixel i in zu, gbm represents the maximum gray value of pixel in zu; N1 represents the number of local regions in FU adjacent to zu, N2 represents the number of local regions in ZU adjacent to zu; α represents the weight;

[0170] Determine Value zu Whether the comparison value is greater than the set judgment value, if so, it means that zu needs to be further divided, if not, it means that zu does not need to be further divided.

[0171] The judgment value of the present invention is calculated from the degree of change of the grayscale value of the pixel points in the local area and the number of local areas around zu that do not belong to FU. The greater the degree of change of the grayscale value of the pixel points in the local area, the greater the number of local areas around zu that do not belong to FU, and the lower the probability that zu needs to be further divided. In this way, the degree of change of the grayscale value of the pixel points in the final area to be corrected can be small, while improving the accuracy of the subsequent correction results, effectively suppressing the number of the final area to be corrected, and realizing the improvement of correction efficiency.

[0172] Preferably, the value of α may be 0.6.

[0173] Preferably, the contrast value of the judgment value is set to 0.3.

[0174] Preferably, each area to be corrected is corrected respectively to obtain a corrected area, including:

[0175] A histogram equalization algorithm is used to correct each area to be corrected to obtain a corrected area.

[0176] Preferably, the pre-processed oblique images are stitched together to obtain a stitched image, comprising:

[0177] Use the following algorithm to stitch images and obtain a stitched image:

[0178] Perform feature point detection on the two images p1 and p2 to be stitched, and obtain the feature points in p1 and p2;

[0179] Use the matching algorithm to obtain the same feature points in p1 and p2 to obtain multiple feature point pairs;

[0180] Eliminate erroneous feature point pairs from multiple feature point pairs to obtain feature point pairs for image registration;

[0181] Estimate the geometric transformation matrix between p1 and p2 based on the feature point pairs used for image registration;

[0182] Based on the geometric transformation matrix, p1 is transformed to the coordinate system where p2 is located to obtain image p3. In the coordinate system where p2 is located, the grayscale values ​​of the pixels in p2 and p3 are weighted averaged to obtain the grayscale values ​​of the pixels in the spliced ​​image.

[0183] Preferably, performing weighted average calculation on the grayscale values ​​of the pixels in p2 and p3 to obtain the grayscale values ​​of the pixels in the spliced ​​image includes:

[0184] For the pixel with coordinates (x, y) in the stitched image, its gray value is represented as G(x, y). If there are pixels with coordinates (x, y) in both p2 and p3, then:

[0185] G(x,y)=w1×p2(x,y)+w2×p3(x,y)

[0186] p2(x,y) represents the grayscale value of the pixel with coordinates (x,y) in p2, p3(x,y) represents the grayscale value of the pixel with coordinates (x,y) in p3, w1 and w2 represent the first weighting coefficient and the second weighting coefficient respectively.

[0187] If there is a pixel with coordinates (x, y) in p2 and there is no pixel with coordinates (x, y) in p3, then G(x, y) = p2(x, y);

[0188] If there is a pixel with coordinates (x, y) in p3, and there is no pixel with coordinates (x, y) in p2, then G(x, y) = p3(x, y).

[0189] After converting p1 to the coordinate system where p2 is located, some pixels in the image p3 obtained will have the same coordinates as the pixels in p2, so it is necessary to determine the grayscale values ​​of these pixels. The present invention uses the first weighting coefficient and the second weighting coefficient to perform weighted summation on the grayscale values ​​of the pixels corresponding to the same coordinates in p2 and p3, which can make the transition of the boundary of the spliced ​​image more natural and retain more edge information.

[0190] Preferably, the calculation formula of w1 is:

[0191]

[0192] w2=1-w1

[0193] d2 represents the distance between the pixel point with coordinates (x, y) and the edge L3, d3 represents the distance between the pixel point with coordinates (x, y) and the edge L2, p23 represents the set of coordinates of the pixel points in the overlapping area between p2 and p3; pl2 represents the area composed of the pixel points in p2 whose coordinates do not belong to p23; L2 is the dividing line between p23 and pl2; pl3 represents the area composed of the pixel points in p3 whose coordinates do not belong to p23; L3 is the dividing line between p23 and pl3, λ represents the reference ratio, neg2 represents the total number of edge pixels contained in the area composed of the pixel points corresponding to the coordinates in the set p23 in p2; neg3 represents the total number of edge pixels contained in the area composed of the pixel points corresponding to the coordinates in the set p23 in p3.

[0194] When calculating w1, the present invention not only considers the distance between the pixel point with coordinates (x, y) and the edge of the overlapping area, but also considers the total number of edge pixels contained in the area corresponding to the coordinates of the pixel points in the overlapping area in p2. Therefore, the more the total number of edge pixels contained in the area corresponding to the coordinates of the pixel points in the overlapping area in p2, the greater the value of d2, the greater the influence of p2 (x, y) on the weighted result, and more image edge information can be retained for the image after splicing while maintaining a natural edge transition, thereby improving the accuracy of subsequent crack identification.

[0195] Preferably, the value of λ is 0.7.

[0196] Preferably, the matching algorithm includes any one of FLANN and KNN.

[0197] Preferably, removing erroneous feature point pairs from a plurality of feature point pairs to obtain feature point pairs for image registration includes:

[0198] The RANSAC algorithm is used to eliminate erroneous feature point pairs from multiple feature point pairs to obtain feature point pairs for image registration.

[0199] The preferred embodiments of the present invention disclosed above are only used to help illustrate the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to the specific implementation methods described. Obviously, many modifications and changes can be made according to the content of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can understand and use the present invention well. The present invention is limited only by the claims and their full scope and equivalents.< / height> < / width>

Claims

1. A dam crack detection system based on drone technology, characterized in that: Including drones and image processing terminals; The drone is used to photograph the slope of the dam and obtain multiple slope images; The image processing terminal is used to identify the inclined surface image and obtain the detection result; Among them, the slope of the dam was photographed, including: Acquire multiple shooting heights based on the height of the slope of the dam; The drone moves horizontally at each shooting height. During the movement, it maintains the same horizontal distance from the slope of the dam. Every fixed moving distance, it takes a picture of the slope of the dam to obtain an image of the slope.

2. A dam crack detection system based on drone technology according to claim 1, characterized in that: The image processing terminal is a computer, which can establish a connection with the drone via WiFi or a data cable. After the connection is established, the computer can access the captured slope images stored in the storage device of the drone.

3. The dam crack detection system based on drone technology according to claim 1 is characterized in that: Obtain multiple shooting heights based on the height of the slope of the dam, including: Let H represent the height of the slope of the dam; D represents the horizontal distance between the drone and the slope of the dam when taking the photo; N represents the number of shooting heights, and N enables the slope images obtained at adjacent shooting heights to overlap in the vertical direction when the UAV shoots the slope of the dam based on D; The first shooting height is The nth shooting height is 4. The dam crack detection system based on drone technology according to claim 3 is characterized in that: The relationship between the area of ​​the overlapping region and the area of ​​the inclined image is: Sa represents the area of ​​the overlapping region in the oblique surface image, Sb represents the area of ​​the oblique surface image, and h represents the set ratio.

5. The dam crack detection system based on drone technology according to claim 4 is characterized in that: The value of h is 0.

5.

6. The dam crack detection system based on drone technology according to claim 3 is characterized in that: Identify the inclined surface image and obtain the detection results, including: Preprocessing each inclined plane image respectively to obtain a preprocessed inclined plane image; Stitching the preprocessed oblique images to obtain a stitched image; Use a deep learning-based detection algorithm to identify the spliced ​​image and obtain the detection result.

7. The dam crack detection system based on drone technology according to claim 6 is characterized in that: Preprocess the inclined image, including: Performing distortion correction processing on the oblique plane image to obtain a distortion-corrected image; Performing illumination correction processing on the distortion-corrected image to obtain an illumination-corrected image; The illumination correction image is subjected to denoising to obtain a preprocessed oblique image.

8. The dam crack detection system based on drone technology according to claim 6 is characterized in that: The pre-processed oblique images are stitched together to obtain a stitched image, including: Use the following algorithm to stitch images and obtain a stitched image: Perform feature point detection on the two images p1 and p2 to be stitched, and obtain the feature points in p1 and p2; Use the matching algorithm to obtain the same feature points in p1 and p2 to obtain multiple feature point pairs; Eliminate erroneous feature point pairs from multiple feature point pairs to obtain feature point pairs for image registration; Estimate the geometric transformation matrix between p1 and p2 based on the feature point pairs used for image registration; Based on the geometric transformation matrix, p1 is transformed to the coordinate system where p2 is located to obtain image p3. In the coordinate system where p2 is located, the grayscale values ​​of the pixels in p2 and p3 are weighted averaged to obtain the grayscale values ​​of the pixels in the spliced ​​image.

9. The dam crack detection system based on drone technology according to claim 8 is characterized in that: The matching algorithm includes any one of FLANN and KNN.

10. The dam crack detection system based on drone technology according to claim 8 is characterized in that: Eliminate the wrong feature point pairs from multiple feature point pairs to obtain feature point pairs for image registration, including: The RANSAC algorithm is used to eliminate erroneous feature point pairs from multiple feature point pairs to obtain feature point pairs for image registration.

Images