A dam displacement monitoring method based on image processing

Abstract

The application discloses a dam displacement monitoring method based on image processing and belongs to the technical field of dam monitoring, and comprises the following steps: S1, installing a target on a dam, determining the initial position of the target on the dam, and collecting real-time monitoring images of the dam; S2, constructing a global vector of the real-time monitoring images by using low-dimensional vectors corresponding to each pixel point in the real-time monitoring images; S3, generating a feature threshold by using a first category coefficient and a second category coefficient of the real-time monitoring images; S4, screening target existing points by using the feature threshold and the global vector of the real-time monitoring images, and determining whether the dam has displacement according to the target existing points and the initial position of the target on the dam. The application can timely issue an early warning when the dam has a slight displacement through real-time monitoring and accurate displacement judgment, is helpful for relevant departments to quickly take response measures, prevents displacement from further expansion, and thus significantly enhances the safety of the dam.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of dam monitoring, and particularly relates to a dam displacement monitoring method based on image processing. BACKGROUND

[0002] Although traditional monitoring methods (such as total station, GPS and tiltmeter) have been widely used, there are still limitations: 1. The disadvantages of contact monitoring: traditional sensors need to be pre-installed on the dam surface or inside, which is limited by construction conditions, and the stability of the equipment and the reliability of the data may decrease when exposed to humid, corrosive and other harsh environments for a long time; 2. The one-sidedness of point monitoring: relying on discrete point data, it is difficult to fully reflect the overall deformation characteristics of the dam, especially the early warning ability of local abnormal displacement; 3. High cost of manual intervention: regular inspection and data calibration require a lot of manpower, and it is difficult to achieve real-time dynamic monitoring.

[0003] In recent years, non-contact monitoring technology based on computer vision has gradually become a research hotspot. By analyzing dam surface image or video data, it extracts structural deformation information, and has the advantages of full scene coverage, low cost and remote monitoring. However, the existing image processing method still faces challenges, on the one hand, the feature expression ability is limited: traditional methods (such as optical flow method and digital image correlation method) rely on pixel-level gray or texture matching, and the calculation complexity is high, and the detection accuracy of small displacement is insufficient; on the other hand, the threshold setting is subjective: the displacement discrimination relies on manual setting of fixed threshold, and it is difficult to adapt to the dynamic changes under different working conditions. SUMMARY

[0004] The application is proposed to solve the above problems, and provides a dam displacement monitoring method based on image processing.

[0005] The technical scheme of the application is as follows: a dam displacement monitoring method based on image processing comprises the following steps:

[0006] S1, installing a target on the dam, determining the initial position of the target on the dam, and collecting real-time monitoring images of the dam;

[0007] S2, constructing a global vector of the real-time monitoring image by using the low-dimensional vector corresponding to each pixel point in the real-time monitoring image;

[0008] S3, generating a feature threshold by using the first category coefficient and the second category coefficient of the real-time monitoring image;

[0009] S4, screening target existing points by using the feature threshold and the global vector of the real-time monitoring image, and determining whether the dam has displacement according to the target existing points and the initial position of the target on the dam.

[0010] Further, S2 comprises the following sub-steps:

[0011] S21. Map the pixel values ​​of each pixel in the real-time monitoring image to a low-dimensional vector;

[0012] S22. Construct a difference function for each pixel;

[0013] S23. Based on the difference function, perform pooling on the low-dimensional vector of the pixel to generate the global vector of the real-time monitoring image.

[0014] The beneficial effects of the above-mentioned further scheme are as follows: In this invention, a 1×1 convolution kernel is used for vector mapping processing to perform linear transformation on pixel values, compressing the high-dimensional pixel value space into a low-dimensional vector space while preserving the relative relationships between pixel values. Through pooling processing, the spatial dimension of the image can be significantly reduced, while also preserving the global information of the image, such as overall shape and texture. The resulting global vector integrates the feature information of all pixels, more comprehensively reflecting the overall state of the dam and providing strong support for subsequent dam displacement determination.

[0015] Furthermore, in S22, the cosine similarity between the low-dimensional vector of a pixel and the low-dimensional vectors of its four neighboring pixels is calculated. The average of the four cosine similarities is taken as the difference factor, and a difference function for the pixel is constructed based on the difference factor. Its expression is:

[0016] ;

[0017] In the formula, A low-dimensional vector representing a pixel. This represents the difference factor of a pixel.

[0018] The beneficial effects of the above-mentioned further scheme are as follows: In this invention, the difference function is used to quantify the difference between a pixel and its neighborhood. By taking the average of the four cosine similarities as the difference factor, the local differences between the pixel and its neighborhood can be captured, resulting in a more stable difference factor. This effectively smooths out noise and improves the robustness of the difference function.

[0019] Furthermore, in S23, feature transformation is performed using the low-dimensional vectors of the corresponding pixels from the activation function, and the feature transformation results of all pixels are input into the pooling function to generate a global vector. Its expression is:

[0020] ;

[0021] In the formula, Indicates average pooling. Indicates the first The difference function of each pixel Indicates the first a weight vector corresponding to the pixel point, a bias vector corresponding to the pixel point, a bias vector corresponding to the pixel point, representing the number of pixel points of the real-time monitoring image.

[0022] The beneficial effects of the above further scheme are: in the present application, the pooling processing can reduce the spatial dimension of all vectors of the image while retaining important feature information, the low-dimensional vector of each pixel point is processed by a feature conversion processing using an activation function, and then the feature conversion results of all pixel points are input to a pooling function to generate a global vector of the real-time monitoring image, and the activation function (such as ReLU) can increase the nonlinearity and improve the expression ability of the global vector.

[0023] Further, S3 comprises the following sub-steps:

[0024] S31, using a clustering algorithm to divide all pixel points of the real-time monitoring image into two categories to obtain a first category set and a second category set;

[0025] S32, extracting the pixel point corresponding to the maximum low-dimensional vector value in the first category set, calculating the cosine similarity between the low-dimensional vector of the pixel point and the low-dimensional vector of each pixel point in the second category set, and adding the square values of all cosine similarities as a first category coefficient;

[0026] S33, extracting the pixel point corresponding to the maximum low-dimensional vector value in the second category set, calculating the cosine similarity between the low-dimensional vector of the pixel point and the low-dimensional vector of each pixel point in the first category set, and adding the square values of all cosine similarities as a second category coefficient;

[0027] S34, calculating the feature threshold of the real-time monitoring image according to the first category coefficient and the second category coefficient.

[0028] The beneficial effects of the above further scheme are: in the present application, the clustering algorithm is an unsupervised learning method, which can divide data points into different categories according to similarity, and divide all pixel points of the real-time monitoring image into two categories, i.e. the first category set and the second category set, to distinguish the pixel points with different features in the image. By calculating the category coefficient, the difference between the two categories can be quantified, and the pixel point corresponding to the maximum low-dimensional vector value is used for calculation, which can highlight the dominant direction influence in this category, so that the category coefficient is more representative.

[0029] Further, in S34, the feature threshold of the real-time monitoring image The calculation formula is:

[0030] ;

[0031] wherein, represents the first category coefficient, represents the second category coefficient, represents the number of pixel points of the real-time monitoring image, represents the number of pixel points of the first category set.

[0032] The above further scheme has the beneficial effect that: in the present application, the feature threshold is dynamically calculated according to the actual situation of the image, which can adapt to the features of different images, and improves the accuracy and robustness of monitoring.

[0033] Further, S4 comprises the following sub-steps:

[0034] S41, generating feature coefficients for each pixel point using the global vector of the real-time monitoring image;

[0035] S42, taking all pixel points with feature coefficients less than the feature threshold as target existence points of the real-time monitoring image;

[0036] S43, determining the actual position of the target on the dam according to the position of the target existence point in the real-time monitoring image;

[0037] S44, judging whether the actual position of the target on the dam is consistent with the initial position of the target on the dam, if yes, the dam has not been displaced, otherwise the dam has been displaced.

[0038] The above further scheme has the beneficial effect that: in the present application, the target region usually has different features from the background region, so its pixel value and global vector value ratio (i.e. feature coefficient) will also be different from the background region. By setting a suitable feature threshold, the target region can be highlighted, which is convenient for subsequent screening. Since the installation position of the target on the dam is known when the image is collected, the position of the target in the image can be converted to the actual position on the dam through the mapping relationship between the image coordinates and the actual coordinates of the dam.

[0039] Further, in S41, the ratio between the pixel value of each pixel point and the global vector value of the real-time monitoring image is taken as the feature coefficient of the pixel point.

[0040] The above further scheme has the beneficial effect that: in the present application, the feature coefficient is the ratio of the pixel value of the pixel point to the global vector value, which reflects the difference between the pixel point and the global image feature. The global vector represents the average feature of the whole image, while the pixel value of the pixel point reflects the local feature. By calculating the ratio of the pixel value to the global vector value, the specificity of the pixel point relative to the global image can be quantified, and this specificity is related to the existence of the target.

[0041] The beneficial effects of the present application are:

[0042] (1) The dam displacement monitoring method provided by the present application realizes a non-contact monitoring mode by installing a target on the dam and collecting real-time monitoring images, avoids the interference on the dam structure caused by the traditional contact monitoring, and reduces the complexity of monitoring equipment maintenance;

[0043] (2) The present application constructs a global vector by using the low-dimensional vector corresponding to each pixel point in the real-time monitoring image, effectively reduces the dimension of data processing, and the global vector can comprehensively reflect the overall deformation of the dam and retain key pixel information;

[0044] (3) The present application divides the pixel points into different categories by using a clustering algorithm, and calculates an adaptive feature threshold based on the category coefficient, avoids the subjectivity and limitation of manually setting a fixed threshold, so that the feature threshold can be dynamically adjusted according to different image features, thereby more accurately judging whether the dam has displacement, and improving the accuracy and robustness of monitoring;

[0045] (4) The present application uses the feature threshold and the global vector to screen the target existence points, and compares with the initial position of the target on the dam, realizes accurate positioning of the target position, provides a reliable basis for judging whether the dam has displacement, and reduces the possibility of misjudgment and omission;

[0046] (5) The present application can timely issue a warning when the dam has a slight displacement through real-time monitoring and accurate displacement judgment, which helps the relevant departments to take prompt measures to prevent the displacement from further expanding, thereby significantly enhancing the safety of the dam. BRIEF DESCRIPTION OF DRAWINGS

[0047] Figure 1 The flowchart of the dam displacement monitoring method based on image processing. DETAILED DESCRIPTION

[0048] The embodiments of the present application will be further described below with reference to the accompanying drawings.

[0049] As shown in the drawings, the present application provides a dam displacement monitoring method based on image processing, comprising the following steps: Figure 1

[0050] S1, installing a target on the dam, determining the initial position of the target on the dam, and collecting real-time monitoring images of the dam;

[0051] S2, constructing a global vector of the real-time monitoring image by using the low-dimensional vector corresponding to each pixel point in the real-time monitoring image;

[0052] S3, generating a feature threshold by using the first category coefficient and the second category coefficient of the real-time monitoring image;

[0053] ​S4, screening target existing points by using feature threshold and global vector of real-time monitoring image, and determining whether the dam has displacement according to target existing points and initial position of target in the dam.

[0054] A target is fixedly installed at a specified position of the dam body (the shape of the target is not limited, and a high-contrast circular / square target can be selected), and initial three-dimensional coordinates of the target are accurately recorded by a professional measuring device (such as a total station), as a reference for subsequent displacement judgment; the target installation needs to ensure firm adhesion to the dam body to avoid self-loosening affecting the monitoring result.

[0055] In the embodiment of the application, S2 comprises the following sub-steps:

[0056] S21, mapping pixel values of each pixel point in the real-time monitoring image into a low-dimensional vector;

[0057] S22, constructing a difference function for each pixel point;

[0058] S23, performing pooling processing on the low-dimensional vector of the pixel point according to the difference function, to generate a global vector of the real-time monitoring image.

[0059] In the application, a 1×1 convolution kernel is used for vector mapping processing, a linear transformation is performed on the pixel value, the high-dimensional pixel value space is compressed to a low-dimensional vector space, and the relative relationship between the pixel values is retained. Through the pooling processing, the spatial dimension of the image can be significantly reduced, and the global information of the image, such as the overall shape and texture, can be retained. The global vector obtained by this way comprehensively reflects the feature information of all pixel points, and can more comprehensively reflect the overall state of the dam, thereby providing strong support for subsequent dam displacement judgment.

[0060] For the collected RGB color image (three-channel pixel value 0-255), a 1×1 convolution kernel is used to perform linear transformation on the RGB value of each pixel point, and the 3-dimensional high-dimensional pixel value is compressed into a 12-dimensional low-dimensional vector; the convolution kernel is pre-trained by gradient descent method, and the initial value is a random normal distribution, so that the relative relationship between the pixels is retained after the mapping.

[0061] For each pixel point, the cosine similarity (cosθ=dot product of two vectors / (product of two vector lengths)) of the low-dimensional vector of the pixel point and the low-dimensional vectors of the four neighboring pixels (up, down and left and right) is calculated, and the average of the four cosine similarities is taken as the difference factor; then the difference function C (12 dimensions) of the pixel is constructed according to the formula to realize local difference quantization and noise smoothing.

[0062] A ReLU activation function is used to convert the feature of the difference function of each pixel (nonlinearity is introduced to avoid feature loss in linear transformation), and the difference function of the mth pixel is , and the corresponding weight vector (12 dimensions, pre-trained fixed), bias vector (12 dimensions), calculated first (σ is the ReLU function);The conversion results of all M pixels are summed and input into the MEANPOOLIING2D (2D average pooling) function, and finally a 12-dimensional real-time monitoring image global vector is generated, which comprehensively reflects the overall features of the image.

[0063] In the embodiment of the application, in S22, the cosine similarity between the low-dimensional vector of the pixel point and the low-dimensional vectors of its four neighborhood pixel points is calculated, the average of the four cosine similarities is taken as the difference factor, and the difference function of the pixel point is constructed according to the difference factor , and its expression is:

[0064] ;

[0065] In the formula, represents the low-dimensional vector of the pixel point, represents the difference factor of the pixel point.

[0066] In the application, the difference function is used to quantify the difference between the pixel point and its neighborhood, and the average of the four cosine similarities is taken as the difference factor, which can capture the local difference between the pixel point and its neighborhood, obtain a more stable difference factor, effectively smooth the noise, and improve the robustness of the difference function.

[0067] In the embodiment of the application, in S23, the low-dimensional vector of the pixel point is processed by using an activation function for feature conversion, and the feature conversion results of all pixel points are input into a pooling function to generate a global vector A, and its expression is:

[0068] ;

[0069] In the formula, represents average pooling, represents the difference function of the i-th pixel point, represents the weight vector corresponding to the i-th pixel point, represents the bias vector corresponding to the i-th pixel point, represents the number of pixel points of the real-time monitoring image.

[0070] ​​​In the present application, the pooling processing can reduce the spatial dimension of all vectors of the image while retaining important feature information, utilize an activation function to perform feature conversion processing on the low-dimensional vector of each pixel point, and then input the feature conversion results of all pixel points into a pooling function to generate a global vector of the real-time monitoring image. The activation function (such as ReLU) can increase nonlinearity and improve the expression ability of the global vector.

[0071] In the embodiment of the present application, S3 comprises the following sub-steps:

[0072] S31, divide all pixel points of the real-time monitoring image into two categories by using a clustering algorithm to obtain a first category set and a second category set;

[0073] S32, extract the pixel point corresponding to the maximum low-dimensional vector value in the first category set, calculate the cosine similarity between the low-dimensional vector of the pixel point and the low-dimensional vectors of all pixel points in the second category set, and add the square values of all cosine similarities to obtain a first category coefficient;

[0074] S33, extract the pixel point corresponding to the maximum low-dimensional vector value in the second category set, calculate the cosine similarity between the low-dimensional vector of the pixel point and the low-dimensional vectors of all pixel points in the first category set, and add the square values of all cosine similarities to obtain a second category coefficient;

[0075] S34, calculate the feature threshold of the real-time monitoring image according to the first category coefficient and the second category coefficient.

[0076] In the present application, the clustering algorithm is an unsupervised learning method that can divide data points into different categories according to similarity, and divide all pixel points of the real-time monitoring image into two categories, i.e., a first category set and a second category set, to distinguish pixel points with different features in the image. By calculating the category coefficient, the difference between the two categories can be quantified, and the pixel point corresponding to the maximum low-dimensional vector value can be used for calculation to highlight the dominant direction influence in the category, making the category coefficient more representative.

[0077] The K-Means clustering algorithm (K=2, unsupervised learning) is used to divide all pixels of the image into a first category set (H pixels) and a second category set (M-H pixels) using the 12-dimensional low-dimensional vector of each pixel as the clustering feature, to realize preliminary separation of target pixels and background pixels.

[0078] The pixel point with the maximum low-dimensional vector length (representing the dominant feature of the category) is extracted from the first category set, the cosine similarity between its low-dimensional vector and the low-dimensional vectors of all pixels in the second category set is calculated, the square values of all similarities are summed to obtain a first category coefficient ; Similarly, the maximum modulus length vector pixel point is extracted from the second category set, and the cosine similarity square sum of all pixel vectors of the first category set is calculated to obtain the second category coefficient , and the difference between the two types of pixels is quantified. Finally, the feature threshold is calculated to ensure that the threshold is dynamically adjusted according to the actual features of the image without human intervention.

[0079] In the embodiment of the present application, in S34, the calculation formula of the feature threshold y of the real-time monitoring image is:

[0080] ;

[0081] In the formula, represents the first category coefficient, represents the second category coefficient, represents the number of pixel points of the real-time monitoring image, represents the number of pixel points of the first category set.

[0082] In the present application, the feature threshold is dynamically calculated according to the actual situation of the image, which can adapt to the features of different images, and improves the accuracy and robustness of the monitoring.

[0083] In the embodiment of the present application, S4 includes the following sub-steps:

[0084] S41, using the global vector of the real-time monitoring image to generate feature coefficients for each pixel point;

[0085] S42, all pixel points with feature coefficients less than the feature threshold are taken as target existence points of the real-time monitoring image;

[0086] S43, according to the position of the target existence point in the real-time monitoring image, the actual position of the target in the dam is determined;

[0087] S44, judging whether the actual position of the target in the dam is consistent with the initial position of the target in the dam, if yes, the dam does not displace, otherwise the dam displaces.

[0088] In the present application, the target region usually has different features from the background region, so the ratio of its pixel value to the global vector value (i.e. the feature coefficient) is also different from the background region. By setting a suitable feature threshold, the target region can be highlighted, which is convenient for subsequent screening. Since the installation position of the target on the dam is known when the image is collected, the position of the target in the image can be converted to the actual position on the dam through the mapping relationship between the image coordinates and the actual coordinates of the dam.

[0089] Calculate the feature coefficients of each pixel, which is the ratio of the mean of the RGB channels of the pixel (pixel value) to the magnitude of the global vector A, to quantify the degree of difference between the pixel and the global image features (the ratio is usually smaller for target pixels due to their special features).

[0090] Pixels with feature coefficients less than the feature threshold are identified as target points, and background interference pixels are removed. Morphological processing (erosion-dilation) is performed on the selected target points to remove isolated noise points and optimize the target region contour.

[0091] The target's point region is fitted using the minimum bounding rectangle, and the center of the rectangle is taken as the target's coordinates (u, v) in the image. Combined with camera calibration parameters (intrinsic parameters: focal length, principal point coordinates; extrinsic parameters: rotation matrix, translation vector), the image coordinates (u, v) are converted into the actual three-dimensional coordinates of the dam. ).

[0092] Displacement determination: Compare with the actual coordinates of the target ( ) and initial coordinates ( ), calculate the displacement in each direction , , If the displacement in any direction is greater than or equal to the preset safety threshold (e.g., 0.5 mm, which can be adjusted according to the dam grade), the dam is determined to have shifted, triggering an early warning (audio-visual alarm + SMS notification).

[0093] In this embodiment of the invention, in S41, the ratio between the pixel value of each pixel and the global vector value of the real-time monitoring image is used as the feature coefficient of the pixel.

[0094] In this invention, the feature coefficient is the ratio of a pixel's pixel value to the global vector value, reflecting the degree of difference between the pixel's features and the global image features. The global vector represents the average features of the entire image, while the pixel value reflects local features. By calculating the ratio of the pixel value to the global vector value, the specificity of a pixel relative to the global image can be quantified; this specificity is related to the presence of the target.

[0095] In this embodiment of the invention, an industrial-grade area scan camera (model optional: Hikvision MV-CA050-10GM) is used, with a resolution of 2448×2048 pixels (single pixel size 3.45μm×3.45μm), a frame rate of 15fps (balancing real-time performance and data storage), an 8mm fixed-focus lens (fixed focal length to avoid image distortion caused by zoom), and an automatic aperture (adapting to changes in lighting conditions from morning to evening, with an aperture range of F1.4-F16).

[0096] Installation position: Install the camera on the fixed observation pier on the downstream side of the dam body (the observation pier is made of C30 concrete with a foundation depth of ≥1.5 m to avoid the influence of foundation settlement on the camera position), the optical axis of the camera is perpendicular to the surface of the dam body with a deviation of ≤3°, and the shooting distance is controlled at 150-200 m (to ensure that the target occupies a range of 15-20 pixels in the image, meeting the identification accuracy).

[0097] Image transmission and storage: The camera transmits real-time images to the background server (configured with Intel Xeon E5-2680 v4 processor, 64 GB memory, and 4 TB SSD hard disk) through gigabit Ethernet (TCP / IP protocol), the server stores image data in real time (storage format is BMP, retains original pixel information, avoids compression distortion), and marks the collection timestamp (accurate to milliseconds). For the collected color images (RGB three channels, each channel pixel value 0-255), a 1x1 convolution kernel is used to linearly transform the RGB value of each pixel point, mapping the 3-dimensional pixel value to a 12-dimensional low-dimensional vector. The ReLU activation function is used to convert the feature of each pixel difference function, the purpose is to introduce nonlinearity and enhance the feature expression ability (to avoid feature loss caused by linear transformation). Let the difference function of the mth pixel be Cm (12-dimensional), the corresponding weight vector wm (12-dimensional, pre-trained by gradient descent method, initial value is random normal distribution), and the bias vector (bm (12-dimensional), the global vector of the real-time monitoring image is obtained.

[0098] Then the K-Means clustering algorithm (K=2, clearly divided into two categories) is used to cluster the features of each pixel 12-dimensional low-dimensional vector, so as to determine the feature threshold. Then the pixels with comprehensive feature coefficient < feature threshold are judged as target existing points.

[0099] Those skilled in the art will appreciate that the embodiments described herein are intended to help the reader understand the principles of the present application and should be understood as not limiting the scope of protection of the present application to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations according to the technical inspiration disclosed in the present application without departing from the essence of the present application, and these modifications and combinations are still within the scope of protection of the present application.

Claims

1. A dam displacement monitoring method based on image processing, characterized in that, Includes the following steps: S1. Install targets on the dam, determine the initial position of the targets on the dam, and collect real-time monitoring images of the dam; S2. Construct a global vector for the real-time monitoring image using the low-dimensional vectors corresponding to each pixel in the real-time monitoring image. S3. Generate feature thresholds using the first and second category coefficients of the real-time monitoring images; S4. Use the feature threshold and global vector of the real-time monitoring image to filter the target existence points, and determine whether the dam has shifted based on the target existence points and the initial position of the target on the dam. S2 includes the following sub-steps: S21. Map the pixel values ​​of each pixel in the real-time monitoring image to a low-dimensional vector; S22. Construct a difference function for each pixel; S23. Based on the difference function, perform pooling on the low-dimensional vector of the pixel to generate the global vector of the real-time monitoring image; In step S22, the cosine similarity between the low-dimensional vector of a pixel and the low-dimensional vectors of its four neighboring pixels is calculated. The average of the four cosine similarities is taken as the difference factor, and a difference function for the pixel is constructed based on the difference factor. Its expression is: ； In the formula, A low-dimensional vector representing a pixel. The difference factor represents the pixel value; In step S23, feature transformation is performed using the low-dimensional vectors of the corresponding pixels from the activation function, and the feature transformation results of all pixels are input into the pooling function to generate a global vector. Its expression is: ； In the formula, Indicates average pooling. Indicates the first The difference function of each pixel Indicates the first The weight vector corresponding to each pixel. Indicates the first The bias vector corresponding to each pixel. σ represents the number of pixels in the real-time monitored image, and σ is the ReLU function.

2. The dam displacement monitoring method based on image processing according to claim 1, characterized in that, S3 includes the following sub-steps: S31. Use a clustering algorithm to divide all pixels of the real-time monitoring image into two categories to obtain the first category set and the second category set; S32. Extract the pixel corresponding to the largest low-dimensional vector value in the first category set, calculate the cosine similarity between the low-dimensional vector of the pixel and the low-dimensional vectors of each pixel in the second category set, and sum the squares of all cosine similarities as the first category coefficient. S33. Extract the pixel corresponding to the largest low-dimensional vector value in the second category set, calculate the cosine similarity between the low-dimensional vector of the pixel and the low-dimensional vectors of each pixel in the first category set, and sum the squares of all cosine similarities as the second category coefficient. S34. Calculate the feature threshold of the real-time monitoring image based on the first category coefficient and the second category coefficient.

3. The dam displacement monitoring method based on image processing according to claim 2, characterized in that, In S34, the feature threshold of the real-time monitoring image The calculation formula is: ； In the formula, Represents the first category coefficient. This represents the second category coefficient. This indicates the number of pixels in the real-time monitored image. This represents the number of pixels in the first category set.

4. The dam displacement monitoring method based on image processing according to claim 1, characterized in that, S4 includes the following sub-steps: S41. Use the global vector of the real-time monitored image to generate feature coefficients for each pixel. S42. All pixels with feature coefficients less than the feature threshold are taken as target points in the real-time monitoring image; S43. Determine the actual location of the target on the dam based on the position of the target's location in the real-time monitoring image; S44. Determine whether the actual position of the target on the dam is consistent with the initial position of the target on the dam. If so, the dam has not shifted; otherwise, the dam has shifted.

5. The dam displacement monitoring method based on image processing according to claim 4, characterized in that, In step S41, the ratio between the pixel value of each pixel and the global vector value of the real-time monitored image is used as the feature coefficient of the pixel.

Images