A dam deformation monitoring method and system

Abstract

The application discloses a dam deformation monitoring method and system, and belongs to the technical field of image processing. Firstly, the Harris corner point response value of each pixel point in a dam remote sensing real-time image and a standard image is calculated, and the pixel points with significant response values in each image region are marked as significant corner points. Then, the image regions with deformation are screened out. Subsequently, the significant corner points of the deformation regions and the corresponding standard image regions are extracted, the corner point distribution real-time image and the standard image are obtained, the added corner point distribution real-time image and the disappeared corner point distribution real-time image are obtained according to the two images, and then the added corner point density and the disappeared corner point density are calculated. Finally, the two kinds of corner point distribution real-time images are processed by using a double-channel dam deformation monitoring neural network, the feature is enhanced by combining the corner point density, and finally the deformation degree value is obtained. Through the accurate corner point analysis and the neural network processing, the precision of the dam deformation monitoring is effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a dam deformation monitoring method and system. Background Art

[0002] Existing image-based dam monitoring technologies fall into two main categories. One approach relies on image grayscale difference calculations. These methods calculate grayscale differences between corresponding regions in remote sensing images of the dam at different times and identify areas where grayscale changes exceed a set threshold as potentially deformed. These methods rely directly on the grayscale information of image pixels. However, the grayscale of a dam's surface is susceptible to factors such as weather (such as rain, fog, and haze) and time of day (such as changes in light angle during sunrise and sunset). Significant grayscale differences can occur even in areas without deformation, leading to the misidentification of numerous falsely deformed areas. The other approach, based on early deep learning, directly predicts deformation from images by constructing neural networks. However, most models rely solely on overall image grayscale information or simple edge features, making it difficult to capture the specific deformation of the dam and effectively distinguishing between deformation and loss, which in turn affects the accuracy of deformation assessment. Summary of the Invention

[0003] In view of the above-mentioned deficiencies in the prior art, the present invention provides a dam deformation monitoring method and system that solves the problem of low accuracy in dam deformation degree assessment in the prior art.

[0004] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is: a dam deformation monitoring method, comprising:

[0005] Harris corner response values ​​are calculated for each pixel in the dam remote sensing real-time image and the dam remote sensing standard image respectively, and pixels with significant Harris corner response values ​​in each image region of the two images are marked as significant corners;

[0006] Calculate the shape feature value of each significant corner point in the image area and filter out the image areas with deformation;

[0007] Extract significant corner points from the image area with deformation and the corresponding image area in the dam remote sensing standard image, and obtain a real-time corner point distribution map and a standard corner point distribution map;

[0008] According to the corner point distribution real-time map and the corner point distribution standard map, a real-time map of added corner point distribution and a real-time map of disappeared corner point distribution are obtained;

[0009] Calculate the density of added corner points and the density of disappeared corner points based on the real-time distribution graph of added corner points and the real-time distribution graph of disappeared corner points;

[0010] A dual-channel dam deformation monitoring neural network is used to process the real-time distribution map of increased corner points and the real-time distribution map of disappeared corner points. Feature enhancement is performed based on the increased corner point density and the disappeared corner point density to obtain the deformation degree value.

[0011] Furthermore, the process of marking significant corners includes: calculating Harris corner response values ​​for each pixel in the dam remote sensing real-time image and the dam remote sensing standard image respectively;

[0012] The dam remote sensing real-time image and the dam remote sensing standard image are divided into multiple image areas respectively;

[0013] Pixels in the image region whose Harris corner response values ​​are greater than the Harris corner response threshold are marked as salient corners.

[0014] Furthermore, the process of screening image regions with deformation includes:

[0015] Calculate the geometric center of salient corners in each image region;

[0016] Construct the corner point covariance matrix based on the geometric center and each significant corner point;

[0017] Solve the corner point covariance matrix to obtain the first and second eigenvalues;

[0018] Calculate the shape eigenvalue according to the first eigenvalue and the second eigenvalue, and construct a shape eigenvector;

[0019] Calculate the similarity between the constructed shape feature vector and the stored shape feature vector;

[0020] When the similarity is less than the similarity threshold, the corresponding image area is deformed.

[0021] Furthermore, the shape feature values ​​include: morphological energy, primary discrete amplitude and secondary discrete amplitude;

[0022] The morphological energy is the product of the first eigenvalue and the second eigenvalue;

[0023] The main discrete amplitude is: the square root of the first eigenvalue;

[0024] The secondary discrete amplitude is: the square root of the second eigenvalue.

[0025] Furthermore, the process of obtaining the corner point distribution real-time map and the corner point distribution standard map includes: setting the pixel values ​​of significant corner points in the deformed image area to 1 and the remaining pixel values ​​to 0 to obtain the corner point distribution real-time map;

[0026] The image area with the same position as the deformed image area is extracted from the standard remote sensing image of the dam. The pixel values ​​of the significant corner points in the image area are set to 1, and the other pixel values ​​are set to 0 to obtain the standard corner point distribution map.

[0027] Furthermore, the process of obtaining the real-time distribution graph of added corner points and the real-time distribution graph of disappeared corner points includes:

[0028] Subtract the corner point distribution real-time map from the corner point distribution standard map at the same pixel position to obtain the corner point difference;

[0029] When the corner point difference is greater than 0, the corresponding pixel point is classified as an added corner point, the pixel value of the added corner point is set to 1, and the pixel values ​​of other positions are set to 0, and a real-time distribution map of the added corner points is obtained;

[0030] When the corner point difference is less than 0, the corresponding pixel point is classified as a vanishing corner point, the pixel value of the vanishing corner point is set to 1, and the pixel values ​​of other positions are set to 0, to obtain a real-time distribution map of the vanishing corner points.

[0031] Furthermore, the formula for calculating the increased corner point density is: , where ρ + To increase the corner density, C + To increase the number of corner points in the corner point distribution real-time map, N is the number of significant corner points in the corner point distribution standard map;

[0032] The formula for calculating the vanishing corner density is: , where ρ - is the vanishing corner density, C - The number of vanishing corner points in the real-time image is the vanishing corner point distribution.

[0033] Furthermore, the dual-channel dam deformation monitoring neural network includes: an added corner processing channel, a vanishing corner processing channel, a multiplier M1, a multiplier M2, and a fully connected layer;

[0034] The input end of the added corner point processing channel is used to input the added corner point distribution real-time map;

[0035] The input end of the vanishing corner processing channel is used to input the real-time map of the vanishing corner distribution;

[0036] The first input terminal of the multiplier M1 is connected to the output terminal of the increased corner point processing channel, and the second input terminal thereof is used for inputting the increased corner point density;

[0037] The first input terminal of the multiplier M2 is connected to the output terminal of the vanishing corner processing channel, and the second input terminal thereof is used to input the vanishing corner density;

[0038] The input end of the fully connected layer is connected to the output end of the multiplier M1 and the output end of the multiplier M2 respectively, and its output end serves as the output end of the dual-channel dam deformation monitoring neural network.

[0039] Furthermore, the structures of the added corner processing channel and the vanishing corner processing channel each include: a first convolution layer, a second convolution layer, a third convolution layer, a multiplier M3, and a continuity measurement layer;

[0040] The input of the first convolutional layer is connected to the input of the continuity measurement layer and serves as the input of the corner addition processing channel and the corner disappearance processing channel;

[0041] A first input terminal of the multiplier M3 is connected to the output terminal of the first convolutional layer, a second input terminal thereof is connected to the output terminal of the continuity measurement layer, and an output terminal thereof is connected to the input terminal of the second convolutional layer;

[0042] The input end of the third convolutional layer is connected to the output end of the second convolutional layer, and its output end serves as the output end of the corner point addition processing channel and the corner point disappearance processing channel;

[0043] The process of obtaining the continuity measurement map of the added corner points in the continuity measurement layer of the added corner point processing channel includes: taking each added corner point as the center, counting the number of added corner points in the neighborhood, and normalizing the number of added corner points to obtain the continuity of the central added corner point, and setting the continuity of other pixel positions to 0;

[0044] The process of obtaining the vanishing corner continuity measurement map in the continuity measurement layer of the vanishing corner processing channel includes: taking each vanishing corner as the center, counting the number of vanishing corners in the neighborhood, and normalizing the number of vanishing corners to obtain the continuity of the central vanishing corner, and setting the continuity of other pixel positions to 0.

[0045] A dam deformation monitoring system includes: a significant corner point marking unit, a deformed image area screening unit, a corner point distribution acquisition unit, a corner point separation unit, a corner point density acquisition unit and a prediction unit;

[0046] The significant corner marking unit is used to calculate the Harris corner response value for each pixel in the dam remote sensing real-time image and the dam remote sensing standard image, and mark the pixel with significant Harris corner response value in each image area of ​​the two images as a significant corner point;

[0047] The deformed image region screening unit is used to calculate shape feature values ​​for significant corner points of each image region and screen image regions with deformation;

[0048] The corner point distribution acquisition unit is used to extract significant corner points from the image area with deformation and the image area in the corresponding dam remote sensing standard image, respectively, to obtain a real-time corner point distribution map and a standard corner point distribution map;

[0049] The corner point separation unit is used to obtain the real-time map of added corner point distribution and the real-time map of disappeared corner point distribution according to the real-time map of corner point distribution and the standard map of corner point distribution;

[0050] The corner point density acquisition unit is used to calculate the density of added corner points and the density of disappeared corner points according to the real-time distribution map of added corner points and the real-time distribution map of disappeared corner points;

[0051] The prediction unit is used to process the real-time map of increased corner point distribution and the real-time map of disappeared corner point distribution using a dual-channel dam deformation monitoring neural network, perform feature enhancement based on the increased corner point density and the disappeared corner point density, and obtain the deformation degree value.

[0052] The beneficial effects of the present invention are:

[0053] 1. This invention obtains significant corner points from each image region of both the real-time and standard remote sensing images of a dam, capturing key feature points on the dam surface. It then calculates shape feature values ​​for each image region in the real-time remote sensing image, reflecting the morphological characteristics of the region and screening out image regions with deformation. This invention uses the shape features of corner points as the core basis for screening deformed regions, rather than relying directly on image grayscale information. Corner points are more stable than image grayscale information and are less susceptible to grayscale fluctuations caused by factors such as weather (rain, haze) and changes in lighting angle. Even if the dam surface exhibits significant grayscale differences due to environmental factors, as long as the corner point's shape features remain unchanged, it will not be misidentified as a deformed region, significantly reducing the probability of false alarms.

[0054] 2. This invention focuses on changes in corner points: When deformation occurs, corner points appear or disappear in local areas. These changes directly reflect subtle structural displacements. By extracting the differences between real-time corner point distribution maps and standard maps, combined with quantitative analysis of the density of increasing and disappearing corner points, the accuracy of dam deformation assessment can be improved.

[0055] 3. Through the steps of "significant corner point extraction-deformation area screening", the present invention first accurately locks the deformed area and eliminates interference from irrelevant areas. Then, it obtains real-time distribution maps of added corner points and real-time distribution maps of disappeared corner points. The added corner point distribution represents the set of newly added corner points, and the disappeared corner point distribution represents the set of disappeared corner points, effectively distinguishing between new deformation information and lost deformation information. Finally, the dual-channel dam deformation monitoring neural network integrates the spatial characteristics of corner point changes (real-time distribution maps of added corner points and real-time distribution maps of disappeared corner points) with quantitative characteristics (corner point density) to improve the accuracy of dam deformation assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 This is a flow chart of a dam deformation monitoring method;

[0057] Figure 2 This is an expanded diagram of the internal structure of the dual-channel dam deformation monitoring neural network;

[0058] Figure 3 Expanded diagram of the internal structure of the added corner processing channel or the vanishing corner processing channel. DETAILED DESCRIPTION

[0059] The specific embodiments of the present invention are described below to facilitate understanding of the present invention by those skilled in the art. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations utilizing the concepts of the present invention are protected.

[0060] Example 1, as Figure 1 As shown, a dam deformation monitoring method includes:

[0061] Harris corner response values ​​are calculated for each pixel in the dam remote sensing real-time image and the dam remote sensing standard image respectively, and pixels with significant Harris corner response values ​​in each image region of the two images are marked as significant corners;

[0062] Calculate the shape feature value of each significant corner point in the image area and filter out the image areas with deformation;

[0063] Extract significant corner points from the image area with deformation and the corresponding image area in the dam remote sensing standard image, and obtain a real-time corner point distribution map and a standard corner point distribution map;

[0064] According to the corner point distribution real-time map and the corner point distribution standard map, a real-time map of added corner point distribution and a real-time map of disappeared corner point distribution are obtained;

[0065] Calculate the density of added corner points and the density of disappeared corner points based on the real-time distribution graph of added corner points and the real-time distribution graph of disappeared corner points;

[0066] A dual-channel dam deformation monitoring neural network is used to process the real-time distribution map of increased corner points and the real-time distribution map of disappeared corner points. Feature enhancement is performed based on the increased corner point density and the disappeared corner point density to obtain the deformation degree value.

[0067] In this embodiment, the process of marking significant corners includes: calculating Harris corner response values ​​for each pixel in the dam remote sensing real-time image and the dam remote sensing standard image respectively;

[0068] The dam remote sensing real-time image and the dam remote sensing standard image are divided into multiple image areas respectively;

[0069] Pixels in the image region whose Harris corner response values ​​are greater than the Harris corner response threshold are marked as salient corners.

[0070] In this embodiment, a 1024×1024 dam remote sensing real-time image and a dam remote sensing standard image are taken as examples to illustrate the process of dividing the image area. Each image area is 64×64 pixels. In the horizontal direction (width direction), 1024 pixels ÷ 64 pixels / area = 16 areas; in the vertical direction (height direction), 1024 pixels ÷ 64 pixels / area = 16 areas; in total, 16×16 = 256 image areas.

[0071] The present invention extracts significant corner points, which are used to display pixel points with drastic changes in grayscale value and sudden changes in gradient direction (such as the intersection of object edges and turning points of texture). They are stable characteristic marks of the local structure of the image.

[0072] In this embodiment, the process of screening image regions with deformation includes:

[0073] Calculate the geometric center of salient corners in each image region;

[0074] Construct the corner point covariance matrix based on the geometric center and each significant corner point;

[0075] Solve the corner point covariance matrix to obtain the first and second eigenvalues;

[0076] Calculate the shape eigenvalue according to the first eigenvalue and the second eigenvalue, and construct a shape eigenvector;

[0077] Calculate the similarity between the constructed shape feature vector and the stored shape feature vector;

[0078] When the similarity is less than the similarity threshold, the corresponding image area is deformed.

[0079] In this embodiment, the expression of the corner point covariance matrix is: , where M is the corner point covariance matrix, x o is the horizontal coordinate of the geometric center, y o is the vertical coordinate of the geometric center, x k is the horizontal coordinate of the kth significant corner point in the image area, y k is the ordinate of the kth salient corner point in the image region, K is the number of salient corner points in the image region, and k is the number of the salient corner point.

[0080] In this embodiment, the Harris corner response threshold is set to 0.04 to 0.06 times the maximum Harris corner response value of the corresponding image.

[0081] In this embodiment, the similarity threshold is set to 1 / 2, and the specific value can be adjusted.

[0082] In this embodiment, the shape feature values ​​include: morphological energy, primary discrete amplitude, and secondary discrete amplitude;

[0083] The morphological energy is the product of the first eigenvalue and the second eigenvalue;

[0084] The main discrete amplitude is: the square root of the first eigenvalue;

[0085] The secondary discrete amplitude is: the square root of the second eigenvalue.

[0086] Dam deformation often manifests itself as changes in the structural morphology of local areas (such as bulging of a certain section of the dam body and expansion of cracks leading to clustering of corner points), and this change will be directly reflected in the "discrete characteristics" of the corner point distribution: for example, the corner points in the dam body settlement area may cluster along the vertical direction (the main discrete amplitude decreases), and the cracked area may increase the overall discrete strength (morphological energy increases) due to the addition of new corner points.

[0087] In this embodiment, the stored shape feature vectors are pre-stored shape feature vectors of the same image region, or are shape feature vectors of the same image region in a dam remote sensing standard image. The similarity is cosine similarity.

[0088] In this embodiment, the first eigenvalue and the second eigenvalue can be used to form a shape feature vector. However, compared with the shape feature vector composed of morphological energy, primary discrete amplitude and secondary discrete amplitude, it cannot reflect the overall discrete strength, and there is a problem of lower accuracy when screening deformed image areas. Therefore, using a shape feature vector composed of morphological energy, primary discrete amplitude and secondary discrete amplitude is a better choice.

[0089] In this embodiment, the process of obtaining the real-time corner point distribution map and the standard corner point distribution map includes: setting the pixel values ​​of significant corner points in the deformed image region to 1 and the remaining pixel values ​​to 0 to obtain the real-time corner point distribution map;

[0090] The image area with the same position as the deformed image area is extracted from the standard remote sensing image of the dam. The pixel values ​​of the significant corner points in the image area are set to 1, and the other pixel values ​​are set to 0 to obtain the standard corner point distribution map.

[0091] In this embodiment, the process of obtaining the real-time distribution map of added corner points and the real-time distribution map of disappeared corner points includes:

[0092] Subtract the corner point distribution real-time map from the corner point distribution standard map at the same pixel position to obtain the corner point difference;

[0093] When the corner point difference is greater than 0, the corresponding pixel point is classified as an added corner point, the pixel value of the added corner point is set to 1, and the pixel values ​​of other positions are set to 0, and a real-time distribution map of the added corner points is obtained;

[0094] When the corner point difference is less than 0, the corresponding pixel point is classified as a vanishing corner point, the pixel value of the vanishing corner point is set to 1, and the pixel values ​​of other positions are set to 0, to obtain a real-time distribution map of the vanishing corner points.

[0095] The present invention sets the pixel values ​​of significant corner points to 1 and the values ​​of other pixels to 0, extracts the corner point distribution, and then subtracts the real-time corner point distribution map from the standard corner point distribution map at the same pixel position. When the corner point difference is equal to 1, the position is a newly added corner point. When the corner point difference is equal to -1, a corner point disappears at the position.

[0096] In this embodiment, the formula for calculating the increased corner point density is: , where ρ + To increase the corner density, C + To increase the number of corner points in the corner point distribution real-time map, N is the number of significant corner points in the corner point distribution standard map;

[0097] The formula for calculating the vanishing corner density is: , where ρ - is the vanishing corner density, C - The number of vanishing corner points in the real-time image is the vanishing corner point distribution.

[0098] If a region ρ + Close to 1, or even greater than 1, when the corner point density increases, it means that the structural characteristics of the area are almost completely reconstructed (such as large-scale cracking of the dam body), which is a "high-risk deformation"; if ρ - If it is close to 1 (the number of vanishing corner points ≈ the total number of standard corner points), it means that the original structural features are almost gone (such as a section of the dam body collapses), and the risk is also extremely high.

[0099] like Figure 2 As shown in FIG, the dual-channel dam deformation monitoring neural network includes: an added corner processing channel, a vanishing corner processing channel, a multiplier M1, a multiplier M2 and a fully connected layer;

[0100] The input end of the added corner point processing channel is used to input the added corner point distribution real-time map;

[0101] The input end of the vanishing corner processing channel is used to input the real-time map of the vanishing corner distribution;

[0102] The first input terminal of the multiplier M1 is connected to the output terminal of the increased corner point processing channel, and the second input terminal thereof is used for inputting the increased corner point density;

[0103] The first input terminal of the multiplier M2 is connected to the output terminal of the vanishing corner processing channel, and the second input terminal thereof is used to input the vanishing corner density;

[0104] The input end of the fully connected layer is connected to the output end of the multiplier M1 and the output end of the multiplier M2 respectively, and its output end serves as the output end of the dual-channel dam deformation monitoring neural network.

[0105] The present invention extracts added corner point features of a real-time graph of added corner point distribution through an added corner point processing channel, extracts vanishing corner point features of a real-time graph of vanishing corner point distribution through a vanishing corner point processing channel, then uses a multiplier M1 to multiply the added corner point density with the output of the added corner point processing channel to adjust the added corner point features (enhancement or weakening), uses a multiplier M2 to multiply the vanishing corner point density with the output of the vanishing corner point processing channel to adjust the vanishing corner point features (enhancement or weakening), and finally outputs a deformation degree value through a fully connected layer.

[0106] The structure of the corner point addition processing channel and the corner point disappearance processing channel is the same, such as Figure 3 As shown, they all include: a first convolutional layer, a second convolutional layer, a third convolutional layer, a multiplier M3 and a continuity measurement layer;

[0107] The input of the first convolutional layer is connected to the input of the continuity measurement layer and serves as the input of the corner addition processing channel and the corner disappearance processing channel;

[0108] A first input terminal of the multiplier M3 is connected to the output terminal of the first convolutional layer, a second input terminal thereof is connected to the output terminal of the continuity measurement layer, and an output terminal thereof is connected to the input terminal of the second convolutional layer;

[0109] The input end of the third convolutional layer is connected to the output end of the second convolutional layer, and its output end serves as the output end of the corner point addition processing channel and the corner point disappearance processing channel.

[0110] In this embodiment, the convolution kernel size of the first convolution layer is 1×1, and the continuity measurement layer is used to calculate the continuity of each corner point. The output of the first convolution layer is element-wise multiplied by the output of the continuity measurement layer through multiplier M3 to achieve the purpose of highlighting the increase of corner points and the disappearance of the area with many corner points, and then further extract features through the second and third convolution layers.

[0111] In this embodiment, the convolution kernel sizes of the second convolution layer and the third convolution layer are 5×5.

[0112] In this embodiment, the process of obtaining the added corner point continuity measurement map in the continuity measurement layer in the added corner point processing channel includes: taking each added corner point as the center, counting the number of added corner points in the neighborhood, and normalizing the number of added corner points to obtain the continuity of the added corner point in the center, and setting the continuity at other pixel points to 0.

[0113] In this embodiment, the process of obtaining the vanishing corner continuity measurement map in the continuity measurement layer in the vanishing corner processing channel includes: taking each vanishing corner as the center, counting the number of vanishing corners in the neighborhood, and normalizing the number of vanishing corners to obtain the continuity of the central vanishing corner, and setting the continuity of other pixel positions to 0.

[0114] In this embodiment, the number of added corner points in a neighborhood range is counted, and the number of disappeared corner points in the neighborhood range is counted. The size of the neighborhood range is 5×5.

[0115] In this embodiment, the normalization formula is: , where ε is continuity, M is the number of added corner points or vanishing corner points, and R is the normalization parameter. When R is set to 5, there are about 5 added corner points or vanishing corner points within a 25-pixel range. The added corner points and vanishing corner points here are relatively dense, indicating that the area has undergone significant changes.

[0116] Example 2, a dam deformation monitoring system, comprising: a significant corner point marking unit, a deformed image region screening unit, a corner point distribution acquisition unit, a corner point separation unit, a corner point density acquisition unit, and a prediction unit;

[0117] The significant corner marking unit is used to calculate the Harris corner response value for each pixel in the dam remote sensing real-time image and the dam remote sensing standard image, and mark the pixel with significant Harris corner response value in each image area of ​​the two images as a significant corner point;

[0118] The deformed image region screening unit is used to calculate shape feature values ​​for significant corner points of each image region and screen image regions with deformation;

[0119] The corner point distribution acquisition unit is used to extract significant corner points from the image area with deformation and the image area in the corresponding dam remote sensing standard image, respectively, to obtain a real-time corner point distribution map and a standard corner point distribution map;

[0120] The corner point separation unit is used to obtain the real-time map of added corner point distribution and the real-time map of disappeared corner point distribution according to the real-time map of corner point distribution and the standard map of corner point distribution;

[0121] The corner point density acquisition unit is used to calculate the density of added corner points and the density of disappeared corner points according to the real-time distribution map of added corner points and the real-time distribution map of disappeared corner points;

[0122] The prediction unit is used to process the real-time map of increased corner point distribution and the real-time map of disappeared corner point distribution using a dual-channel dam deformation monitoring neural network, perform feature enhancement based on the increased corner point density and the disappeared corner point density, and obtain the deformation degree value.

[0123] The specific implementation process of Example 2 is the same as that of Example 1.

[0124] Table 1

[0125]

[0126] This method obtains significant corner points from each image region of both real-time and standard remote sensing images of a dam, capturing key feature points on the dam surface. It then calculates shape feature values ​​for each image region in the real-time remote sensing image, reflecting the morphological characteristics of the region and screening out image regions with deformation. This method uses the shape features of corner points as the core basis for screening deformed regions, rather than relying directly on image grayscale information. Corner points are more stable than image grayscale information and are less susceptible to grayscale fluctuations caused by factors such as weather (rain, haze) and changes in lighting angle. Even if the dam surface exhibits significant grayscale differences due to environmental factors, as long as the corner point's shape features remain unchanged, it will not be misidentified as a deformed region, significantly reducing the probability of false alarms.

[0127] This method focuses on changes in corner points: When deformation occurs, corner points appear or disappear in a local area. These changes directly reflect subtle structural displacements. By extracting the differences between a real-time corner point distribution map and a standard map, combined with quantitative analysis of the density of increasing and disappearing corner points, the accuracy of dam deformation assessment can be improved.

[0128] The present invention uses the steps of "significant corner point extraction-deformation area screening" to first accurately lock the area where deformation exists and eliminate interference from irrelevant areas. Subsequently, real-time maps of the distribution of added corner points and the distribution of disappeared corner points are obtained. The added corner point distribution represents the set of newly added corner points, and the disappeared corner point distribution represents the set of disappeared corner points, effectively distinguishing between new deformation information and lost deformation information. Finally, the dual-channel dam deformation monitoring neural network fuses the spatial characteristics of corner point changes (real-time maps of the distribution of added corner points and the distribution of disappeared corner points) with quantitative characteristics (corner point density) to improve the accuracy of dam deformation degree assessment.

[0129] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A dam deformation monitoring method, characterized in that: include: Harris corner response values ​​are calculated for each pixel in the dam remote sensing real-time image and the dam remote sensing standard image respectively, and pixels with significant Harris corner response values ​​in each image region of the two images are marked as significant corners; Calculate the shape feature value of each significant corner point in the image area and filter out the image areas with deformation; Extract significant corner points from the image area with deformation and the corresponding image area in the dam remote sensing standard image, and obtain a real-time corner point distribution map and a standard corner point distribution map; According to the corner point distribution real-time map and the corner point distribution standard map, a real-time map of added corner point distribution and a real-time map of disappeared corner point distribution are obtained; Calculate the density of added corner points and the density of disappeared corner points based on the real-time distribution graph of added corner points and the real-time distribution graph of disappeared corner points; A dual-channel dam deformation monitoring neural network is used to process the real-time distribution map of increased corner points and the real-time distribution map of disappeared corner points. Feature enhancement is performed based on the increased corner point density and the disappeared corner point density to obtain the deformation degree value.

2. The dam deformation monitoring method according to claim 1, characterized in that: The process of marking significant corners includes: calculating Harris corner response values ​​for each pixel in the dam remote sensing real-time image and the dam remote sensing standard image; The dam remote sensing real-time image and the dam remote sensing standard image are divided into multiple image areas respectively; Pixels in the image region whose Harris corner response values ​​are greater than the Harris corner response threshold are marked as salient corners.

3. The dam deformation monitoring method according to claim 1, characterized in that: The process of screening image areas for deformation includes: Calculate the geometric center of salient corners in each image region; Construct the corner point covariance matrix based on the geometric center and each significant corner point; Solve the corner point covariance matrix to obtain the first and second eigenvalues; Calculate the shape eigenvalue according to the first eigenvalue and the second eigenvalue, and construct a shape eigenvector; Calculate the similarity between the constructed shape feature vector and the stored shape feature vector; When the similarity is less than the similarity threshold, the corresponding image area is deformed.

4. The dam deformation monitoring method according to claim 3, characterized in that: The shape eigenvalues ​​include: morphological energy, primary discrete amplitude and secondary discrete amplitude; The morphological energy is the product of the first eigenvalue and the second eigenvalue; The main discrete amplitude is: the square root of the first eigenvalue; The secondary discrete amplitude is: the square root of the second eigenvalue.

5. The dam deformation monitoring method according to claim 1, characterized in that: The process of obtaining the corner point distribution real-time map and the corner point distribution standard map includes: setting the pixel values ​​of the significant corner points in the deformed image area to 1 and the other pixel values ​​to 0 to obtain the corner point distribution real-time map; The image area with the same position as the deformed image area is extracted from the standard remote sensing image of the dam. The pixel values ​​of the significant corner points in the image area are set to 1, and the other pixel values ​​are set to 0 to obtain the standard corner point distribution map.

6. The dam deformation monitoring method according to claim 1, characterized in that: The process of obtaining the real-time distribution graph of added corner points and the real-time distribution graph of disappeared corner points includes: Subtract the corner point distribution real-time map from the corner point distribution standard map at the same pixel position to obtain the corner point difference; When the corner point difference is greater than 0, the corresponding pixel point is classified as an added corner point, the pixel value of the added corner point is set to 1, and the pixel values ​​of other positions are set to 0, and a real-time distribution map of the added corner points is obtained; When the corner point difference is less than 0, the corresponding pixel point is classified as a vanishing corner point, the pixel value of the vanishing corner point is set to 1, and the pixel values ​​of other positions are set to 0, to obtain a real-time distribution map of the vanishing corner points.

7. The dam deformation monitoring method according to claim 1, characterized in that: The formula for calculating the increased corner point density is: , where ρ + To increase the corner density, C + To increase the number of corner points in the corner point distribution real-time map, N is the number of significant corner points in the corner point distribution standard map; The formula for calculating the vanishing corner density is: , where ρ - is the vanishing corner density, C - The number of vanishing corner points in the real-time image is the vanishing corner point distribution.

8. The dam deformation monitoring method according to claim 1, characterized in that: The dual-channel dam deformation monitoring neural network includes: an added corner processing channel, a vanishing corner processing channel, a multiplier M1, a multiplier M2 and a fully connected layer; The input end of the added corner point processing channel is used to input the added corner point distribution real-time map; The input end of the vanishing corner processing channel is used to input the real-time map of the vanishing corner distribution; The first input terminal of the multiplier M1 is connected to the output terminal of the increased corner point processing channel, and the second input terminal thereof is used for inputting the increased corner point density; The first input terminal of the multiplier M2 is connected to the output terminal of the vanishing corner processing channel, and the second input terminal thereof is used to input the vanishing corner density; The input end of the fully connected layer is connected to the output end of the multiplier M1 and the output end of the multiplier M2 respectively, and its output end serves as the output end of the dual-channel dam deformation monitoring neural network.

9. The dam deformation monitoring method according to claim 8, characterized in that: The structures of the added corner processing channel and the vanishing corner processing channel both include: a first convolution layer, a second convolution layer, a third convolution layer, a multiplier M3 and a continuity measurement layer; The input of the first convolutional layer is connected to the input of the continuity measurement layer and serves as the input of the corner addition processing channel and the corner disappearance processing channel; A first input terminal of the multiplier M3 is connected to the output terminal of the first convolutional layer, a second input terminal thereof is connected to the output terminal of the continuity measurement layer, and an output terminal thereof is connected to the input terminal of the second convolutional layer; The input end of the third convolutional layer is connected to the output end of the second convolutional layer, and its output end serves as the output end of the corner point addition processing channel and the corner point disappearance processing channel; The process of obtaining the continuity measurement map of the added corner points in the continuity measurement layer of the added corner point processing channel includes: taking each added corner point as the center, counting the number of added corner points in the neighborhood, and normalizing the number of added corner points to obtain the continuity of the central added corner point, and setting the continuity of other pixel positions to 0; The process of obtaining the vanishing corner continuity measurement map in the continuity measurement layer of the vanishing corner processing channel includes: taking each vanishing corner as the center, counting the number of vanishing corners in the neighborhood, and normalizing the number of vanishing corners to obtain the continuity of the central vanishing corner, and setting the continuity of other pixel positions to 0.

10. A dam deformation monitoring system, implemented based on the dam deformation monitoring method according to any one of claims 1 to 9, characterized in that: include: Significant corner marking unit, deformed image area screening unit, corner distribution acquisition unit, corner separation unit, corner density acquisition unit and prediction unit; The significant corner marking unit is used to calculate the Harris corner response value for each pixel in the dam remote sensing real-time image and the dam remote sensing standard image, and mark the pixel with significant Harris corner response value in each image area of ​​the two images as a significant corner point; The deformed image region screening unit is used to calculate shape feature values ​​for significant corner points of each image region and screen image regions with deformation; The corner point distribution acquisition unit is used to extract significant corner points from the image area with deformation and the image area in the corresponding dam remote sensing standard image, respectively, to obtain a real-time corner point distribution map and a standard corner point distribution map; The corner point separation unit is used to obtain the real-time map of added corner point distribution and the real-time map of disappeared corner point distribution according to the real-time map of corner point distribution and the standard map of corner point distribution; The corner point density acquisition unit is used to calculate the density of added corner points and the density of disappeared corner points according to the real-time distribution map of added corner points and the real-time distribution map of disappeared corner points; The prediction unit is used to process the real-time map of increased corner point distribution and the real-time map of disappeared corner point distribution using a dual-channel dam deformation monitoring neural network, perform feature enhancement based on the increased corner point density and the disappeared corner point density, and obtain the deformation degree value.

Images